An artificial intelligence correction method for in-orbit installation matrix of remote sensing satellite instruments

Abstract

This invention discloses an artificial intelligence correction method for the on-orbit installation matrix of remote sensing satellite instruments. The method uses a neural network to store the change law of the remote sensing instrument installation matrix as a correction device. Based on artificial intelligence technology, the boundary conditions that determine the state value of the remote sensing instrument installation matrix are used as inputs, and the state value of the remote sensing satellite instrument installation matrix is ​​used as the output to train the device to obtain its parameters. Then, when the installation matrix state is needed, the steady-state temperature field of the satellite body and the instrument is input into the device to obtain the installation matrix state in a timely manner.

Description

Technical Field

[0001] The invention belongs to the field of aerospace technology, and in particular relates to an artificial intelligence correction method for an on-orbit installation matrix of a remote sensing satellite instrument. Background Art

[0002] The instrument installation matrix of a remote sensing satellite refers to a 3×3 rotation matrix between two coordinate systems determined by the instrument installation angle. Its specific composition is shown in Formula 1. The instrument installation angle refers to the angle between the coordinate axis of the instrument coordinate system and the coordinate axis of the satellite body coordinate system. Generally, the instrument installation angle is accurately measured when the satellite is installed and adjusted. However, it is inevitable that the measurement accuracy of the instrument installation angle, the vibration during the satellite launch, and the thermal deformation caused by the thermal expansion and contraction of the satellite after it is in orbit will reduce the accuracy of the instrument installation angle. Therefore, it is necessary and even indispensable to measure the installation angle in orbit. This is especially important for geostationary three-axis stabilized remote sensing satellites, because the characteristics of the geostationary orbit determine that the external heat flux of the sun heats the three-axis stabilized satellite non-uniformly, and the heating is typically carried out in a daily cycle. Although the influence of external heat flux is considered when designing remote sensing instruments, and the size of thermal expansion and contraction in orbit is minimized as much as possible, such as using materials with low thermal expansion and contraction coefficients, the installation angle changes still exist.

[0003] Matrix IM =R z (-ψ)*R x (-θ)*R y (-φ) (1)

[0004] Where ψ, θ, φ are the installation angles, R p (-q), p = x, y, z represents the rotation of -q angle around the P axis. Its exact form can be found in the formula

[0005] Formula 2, Matrix IM It is a 3×3 rotation matrix representing the rotation from the instrument coordinate system to the satellite body coordinate system.

[0006]

[0007] Due to the change in the instrument installation angle, Matrix IM It also changes accordingly, so it is necessary and essential to make on-orbit corrections. Among the existing correction methods, the traditional method of using Fourier expansion method to describe the law of installation angle change over time is commonly used, see formula 3.

[0008]

[0009] The Fourier expansion form of formula 3 is used to describe the law of the installation angle changing with time, and the coefficients are obtained by fitting the observed data to obtain the actual installation angle at any time point. However, Fourier expansion is a mathematical method that does not consider any physical factors and physical processes of changes in the installation matrix. Its accuracy for the object being fitted generally increases with the increase of the adopted order. In other words, there must be errors in fitting using the finite-term Fourier expansion method. Summary of the invention

[0010] In order to solve the problems existing in the prior art, the present invention proposes an artificial intelligence correction method for an on-orbit installation matrix of a remote sensing satellite instrument. A neural network is used to store the variation law of the remote sensing instrument installation matrix as a correction device. Based on artificial intelligence technology, the boundary conditions that determine the state value of the remote sensing instrument installation matrix are used as input, and the state value of the remote sensing satellite instrument installation matrix is ​​used as output. The device is trained to obtain its parameters. Then, when the installation matrix state is required, the steady-state ambient temperature field of the satellite body and the instrument is input into the device, so as to obtain the installation matrix state in time.

[0011] To achieve the above object, the present invention provides an artificial intelligence correction method for an on-orbit installation matrix of a remote sensing satellite instrument, the steps of which are as follows:

[0012] S1, based on the known remote sensing satellite orbit, attitude and remote sensing instrument observation geometry, calculates the actual observation direction of a specific point on the remote sensing image through the position control point;

[0013] S2, using the observation orientation obtained in S1 as a known condition, calculates the state value of the instrument installation matrix on the remote sensing satellite when observing the position control point;

[0014] S3, taking the boundary conditions that affect the remote sensing satellite instrument installation matrix as input and the instrument installation matrix state value in S2 as label, trains the IMNN device that describes the change law of the installation matrix to obtain the neural network weight;

[0015] S4, through measurement, the boundary conditions that affect the remote sensing satellite instrument installation matrix at a specific moment are obtained, and then these boundary conditions are input into the IMNN device with known neural network weights to obtain the state value of the instrument installation matrix at a specific moment, which is used in subsequent remote sensing data positioning or other data processing.

[0016] Furthermore, the boundary condition affecting the remote sensing satellite instrument installation matrix in S3 is the steady-state ambient temperature field when the satellite body and the instrument are in thermal equilibrium.

[0017] Furthermore, the position control point in S1 may be a space control point such as a star or the sun, or a surface control point such as a coastline with known longitude and latitude and other specific targets.

[0018] Furthermore, the state value of the instrument installation matrix on the remote sensing satellite when observing the position control point in S2 is calculated using a gradient descent search algorithm, wherein the search starting condition is to set θ, φ, ψ as the installation matrix measured on the ground before launch.

[0019] Furthermore, the training steps of the IMNN device in S4 are:

[0020] (1) The steady-state ambient temperature field of the satellite body and instruments in thermal equilibrium is used as input, and θ, φ, ψ are used as labels for training;

[0021] (2) During training, all inputs and labels are normalized;

[0022] (3) constructing a fully connected neural network containing an input layer, at least one hidden layer and an output layer to form an IMNN;

[0023] (4) Train the defined IMNN using the normalized input and labels.

[0024] Furthermore, when a sun-synchronous orbit satellite uses this method, the duration of the training data used is limited to be greater than one orbital period, and when a geosynchronous orbit satellite uses this method, the duration of the training data used is limited to be greater than 24 hours.

[0025] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0026] Based on the high-precision fitting ability of neural networks for functions, and considering the fundamental physical factors that cause changes in the installation matrix of remote sensing satellite instruments, this method can efficiently and high-quality output the current true value of the remote sensing satellite instrument installation matrix. When performing remote sensing data positioning or other processing, the true value is substituted into the processing process to eliminate the influence of the change of the remote sensing satellite instrument installation matrix and obtain high-quality positioning or other processing results. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0028] Figure 1 A flow chart of an artificial intelligence correction method for an on-orbit installation matrix of a remote sensing satellite instrument provided in an embodiment of the present application;

[0029] Figure 2 Comparison chart of the Fourier decomposition installation matrix and the neural network fitting installation matrix.

[0030] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] English abbreviations meaning:

[0033] IMNN: Installing Matrix Neural Networks.

[0034] The installation matrix is ​​stored in the parameter file as an important parameter and is repeatedly called by the program. The intersection position of the optical axis of the earth remote sensing satellite and the surface needs to be obtained through precise mathematical conversion of the instrument installation matrix. Experience shows that due to the long-distance observation characteristics of remote sensing satellites, the angle deviation will be magnified by the distance. For example, in the extremely important geostationary orbit (orbit altitude 35,800km from the earth's surface) remote sensing observation, when the satellite-side angle error is 0.0016°, the optical axis has deviated from the accurate position on the surface by 1km, resulting in a significant decrease in the quality of remote sensing data. Therefore, after the satellite is launched, the instrument installation matrix needs to be re-estimated and corrected to calculate the high-precision geographic location of the observation target. If it is not re-estimated, the positioning result will have a large systematic deviation, which will reduce the data quality and affect the application effect.

[0035] Figure 2 The deformation variable y generated by a single variable x under the action of the function as shown in the figure is simulated, and then 96 points are selected on the curve y to represent the deformation variable determined by the control points every 15 minutes in a day (24 hours). The thermal deformation curve y is constructed by Fourier decomposition and artificial intelligence fitting respectively, so that the corresponding y can be output by inputting x at any time.

[0036] For Fourier decomposition, its fitting accuracy is greatly affected by the choice of order and the form of the function to be fitted. The differences in the on-orbit configurations of different satellites make the thermal deformation vary greatly, so strong prior knowledge is required to complete the Fourier decomposition work.

[0037] Theoretically, a neural network can fit a continuous function with high precision, so there is little need for prior knowledge. Therefore, the embodiment of the present invention facilitates any satellite to complete high-precision thermal deformation fitting.

[0038] Specifically, when the variable x that causes thermal deformation y on the actual satellite is greater than 1, since the Fourier decomposition only takes time t as a variable, the physical process of the variable x causing thermal deformation cannot be constructed; while the neural network takes all variables x as network inputs, so the process of the variable causing thermal deformation can be physically described.

[0039] Under the condition of accurate physical model, the neural network fitting method constructs the thermal deformation physical process by itself through training, so thermal deformation fitting and correction can be performed with high precision.

[0040] The following is an example of an embodiment of the present invention to illustrate a correction method for a remote sensing satellite instrument installation matrix artificial intelligence. Figure 1 As shown:

[0041] S1. Based on the known remote sensing satellite orbit, attitude and remote sensing instrument observation geometry, the actual observation direction of a specific point on the remote sensing image (the image where the position control point is observed) is calculated through the position control point (which can be a space control point - a star, the sun, etc., or a surface control point - a coastline with known longitude and latitude, a specific target, etc.).

[0042] As shown in formula 4.

[0043]

[0044] The coordinates of the position control point in the ground-fixed system (or other coordinate systems that can be converted from the ground-fixed system) are (x L ,y L ,z L ), the coordinates of the satellite are (x SAT ,y SAT ,z SAT ), then the unit observation vector can be calculated

[0045] S2. Using the known observation orientation as a known condition, a gradient descent search algorithm is used to calculate the state value of the instrument installation matrix on the remote sensing satellite when observing the position control point (the 3×3 rotation matrix of the instrument coordinate system to the satellite coordinate system).

[0046]

[0047] Formula 5 is the complete rotation process from the observation vector in the instrument coordinate system to the observation vector in the earth-fixed system, and is equal to the observation vector of the earth-fixed position control point calculated by the position control point. Analyzing formula 5, A, B, C, and D are the polar motion matrix, the sidereal time matrix, the nutation matrix, and the precession matrix, respectively. Obviously, when the specific satellite remote sensing instrument observation process is determined, they are all astronomical parameter matrices related to the observation time. Under current astronomical conditions, they can be regarded as known quantities.

[0048] R ORB-ECI It is a function of the six orbital numbers a, e, i, Ω, ω, and M. The six orbital numbers are the output of satellite ranging and orbit determination and have nothing to do with instrument observations. They can also be regarded as known quantities here.

[0049] R SAT-ORB It is a function of the satellite attitude. Currently, remote sensing satellites usually carry attitude measurement equipment such as star sensors, horizon instruments and gyroscopes. It has nothing to do with instrument observations and can be regarded as a known quantity here.

[0050] R INS-SAT It is determined by the installation angle of the remote sensing instrument on the satellite body, which can be called the installation matrix. Practice has shown that the installation matrix will generally change after the satellite enters orbit due to reasons such as installation matrix measurement accuracy and launch vibration. In addition, since the satellite will expand and contract while in orbit, the instrument installation matrix will change over time. This is especially obvious for geostationary satellites. Therefore, it is necessary to determine the installation matrix on orbit for satellites with large deformations in order to determine the vector rotation equation to complete the geometric positioning of the remote sensing image.

[0051] R MIRROR The vector in the instrument coordinate system after the instrument detector output vector is reflected by the scanning mirror, where (ρ1, ρ2, ρ3) is the detector output vector,

[0052] Determined by the structure of the instrument's optical system, the satellite's optical structure can be considered a known quantity after precise measurement on the ground before launch.

[0053] According to the analysis of the rotation matrices in Formula 5 (all 3×3 matrices), let

[0054] W1=A*B*C*D*R ORB-ECI *R SAT-ORB (6)

[0055] So formula 5 can be written as:

[0056]

[0057] Among them, only R INS-SAT is an unknown quantity. In order to find this unknown quantity, further transforming formula 7 can be obtained

[0058]

[0059] in,

[0060] Substituting Equation 1 into Equation 8 yields Equation 9:

[0061]

[0062] According to Formula 2, Formula 9 is a set of three equations with θ, φ, ψ as unknown variables, and θ, φ, ψ are combined as parameters with actual physical meanings.

[0063] Here we use a search algorithm based on gradient descent,

[0064]

[0065] In the formula

[0066] Set the search stop condition ε

[0067] if(F>ε)

[0068]

[0069]

[0070] else

[0071] return(x i )

[0072] end

[0073] The search starting condition is to let θ, φ, ψ be the installation matrix measured on the ground before launch. Based on the fact that the installation angle variation range due to thermal expansion and contraction will not be too large, it is reasonable to let θ, φ, ψ be the installation matrix measured on the ground before launch, which also facilitates the search results to converge quickly to the actual value.

[0074] S3. Taking the boundary conditions that affect the installation matrix of the remote sensing satellite instrument as input and the state values ​​of the instrument installation matrix when observing the position control points solved based on the position control points as labels, the installation matrix neural network (IMNN) that describes the change law of the installation matrix is ​​trained to obtain the network parameters of the IMNN, i.e., the neural network weights.

[0075] After obtaining θ, φ, and ψ through S2, temperature measurement points are set on the satellite body (satellite platform) and the instrument to monitor the on-orbit working status of the satellite. The on-orbit status of the satellite (including the platform and the instrument) is measured. Since the change of the installation matrix is ​​essentially the thermal expansion and contraction of the satellite (including the platform and the instrument) structure, it is in line with objective facts to use the steady-state ambient temperature field when the satellite body and the instrument are in thermal equilibrium as the boundary condition for the change of the installation matrix.

[0076] Since the physical fact that the installation matrix undergoes thermal expansion and contraction determines that the installation matrix changes continuously, it is regarded as a mathematical continuous function, so a fully connected network structure is used to train the IMNN, with the steady-state ambient temperature field when the satellite body and the instrument are in thermal equilibrium as input, and θ, φ, ψ as labels for training. The Adam optimizer can be used (but not limited to). During training, the following steps are performed:

[0077] (1) The steady-state ambient temperature field of the satellite body and instruments in thermal equilibrium is used as input, and θ, φ, ψ are used as labels for training;

[0078] (2) Normalize all inputs and labels,

[0079]

[0080] where t i represents the temperature of the ith temperature measurement point, t i-max and t i-min The installation angle normalization and temperature normalization adopt the same method.

[0081] (3) Then, a fully connected neural network including an input layer, at least one hidden layer and an output layer is constructed using the following code (using Python code as an example, but not limited to Python code) to form an IMNN.

[0082] model=Sequential(

[0083] Linear(input_feature,hidden_feature),

[0084] ActiveFunction,

[0085] Linear(hidden_feature,hidden_feature),

[0086] ActiveFunction,

[0087] Linear(hidden_feature,output_feature),

[0088] ActiveFunction )

[0090] (4) Use the input (normalized temperature) and label (normalized installation angle) to train the defined IMNN. Since the satellite orbit has a certain periodicity, when this method is used for a sun-synchronous orbit satellite, the training data used must be longer than one orbital period, and when this method is used for a geosynchronous orbit satellite, the training data used must be longer than 24 hours.

[0091] The logical process of training is as follows (using Python code as an example, but not limited to Python code). forepoch in tqdm.tqdm(range(1,epochs+1)):

[0092] #Forward propagation

[0093] y_pred = model(x)

[0094] loss = loss_f(y_pred,y)

[0095] Epoch.append(epoch)

[0096] Loss.append(loss.data)

[0097] if epoch%100==0:

[0098] print("Epoch:{},loss:{}".format(epoch,loss))

[0099] optim.zero_grad()

[0100] # Backward propagation

[0101] loss.backward()

[0102] # Parameter fine-tuning

[0103] optim.step()

[0104] The loss function is form.

[0105] S4. The boundary conditions that affect the remote sensing satellite instrument installation matrix at a specific moment are obtained through measurement, and then these boundary conditions are input into the IMNN with known neural network weights. The state value of the instrument installation matrix at a specific moment is obtained through the IMNN network, such as formula 11, which is used in subsequent remote sensing data positioning or other data processing.

[0106] (φ,θ,ψ)=IMNN.out(t i ) (11)

[0107] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiment part.

[0108] The professionals should also be further aware that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0109] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for artificial intelligence correction of on-orbit installation matrix of remote sensing satellite instruments, characterized in that: The method comprises the following steps: S1, based on the known remote sensing satellite orbit, attitude and remote sensing instrument observation geometry, calculates the actual observation direction of the preset point on the remote sensing image through the position control point; S2, using the observation orientation obtained in S1 as a known condition, calculates the state value of the instrument installation matrix on the remote sensing satellite when observing the position control point; S3, taking the boundary conditions that affect the remote sensing satellite instrument installation matrix as input and the instrument installation matrix state value in S2 as label, trains the installation matrix neural network IMNN that describes the change law of the installation matrix to obtain the neural network weights; S4, obtain the boundary conditions that affect the remote sensing satellite instrument installation matrix at a certain moment through measurement, and then input these boundary conditions into the IMNN with known neural network weights to obtain the state value of the instrument installation matrix at that moment for subsequent remote sensing data positioning; The boundary condition affecting the remote sensing satellite instrument installation matrix in S3 is the steady-state ambient temperature field when the satellite body and the instrument are in thermal equilibrium.

2. According to claim 1, a remote sensing satellite instrument on-orbit installation matrix artificial intelligence correction method is characterized in that: The position control points in S1 include space control points such as stars and the sun, and surface control points such as coastlines with known longitude and latitude.

3. The artificial intelligence correction method for in-orbit installation matrix of a remote sensing satellite instrument according to claim 1 is characterized in that: The state value of the instrument installation matrix on the remote sensing satellite when observing the position control point in S2 is calculated using the gradient descent search algorithm, where the search starting condition is It is the installation matrix status value measured on the ground before launch.

4. The method for artificial intelligence correction of on-orbit installation matrix of remote sensing satellite instrument according to claim 3 is characterized in that: The training steps of IMNN in S4 are: (1) Taking the steady-state ambient temperature field when the satellite body and instruments are in thermal equilibrium as input, Training for labels; (2) During training, all inputs and labels are normalized; (3) constructing a fully connected neural network containing an input layer, at least one hidden layer and an output layer to form an IMNN; (4) Train the defined IMNN using the normalized input and labels.

5. The method for artificial intelligence correction of on-orbit installation matrix of remote sensing satellite instrument according to claim 4 is characterized in that: When this method is used by a sun-synchronous orbit satellite, the training data duration used is limited to be greater than one orbital period. When this method is used by a geosynchronous orbit satellite, the training data duration used is limited to be greater than 24 hours.

Images