Space on-orbit maneuvering target monocular tracking algorithm based on variable state dimensions
By employing a monocular tracking algorithm for maneuvering space targets with variable state dimensions, combined with optical observation and extended Kalman filtering, the real-time and accuracy issues under monocular observation conditions are resolved, enabling real-time, accurate, and universal tracking of maneuvering space targets.
Patent Information
- Application Number
- CN202411036292.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-19
AI Technical Summary
Existing space maneuvering target tracking algorithms lack real-time performance and accuracy under monocular observation conditions, and the limited observation equipment makes it difficult to meet the requirements of convenience.
A monocular tracking algorithm for space maneuvering targets based on variable state dimension is adopted. The target is tracked through an optical observation platform. The extended Kalman filter and innovation fading memory averaging are combined for maneuver detection and acceleration estimation. The normal and increased-dimensional state models are switched for real-time tracking.
It achieves real-time and accurate tracking under single-eye optical observation conditions, possesses good tracking capabilities and versatility, and is suitable for the convenience requirements of optical observation platforms.
Smart Images

Figure CN118980376A_ABST
Abstract
Description
Technical Field
[0001] The present invention provides a real-time and accurate space maneuvering target tracking algorithm based on variable state dimension and suitable for monocular observation conditions. The method involves tracking the target using the increased-dimensional state variables composed of the position, velocity and maneuvering acceleration of the space target and the single-line-of-sight observation means based on the optical sensor observation platform, as well as the algorithm of using the test statistic based on the new information to detect the start of maneuvering of the target and the test statistic based on the maneuvering acceleration estimation to detect the end of maneuvering of the target. It belongs to the field of navigation technology. Background Art
[0002] With the increasing number and types of satellites in various countries, as well as the continuous expansion and improvement of satellite application fields and functions, the space application system is becoming more and more perfect. The increasing number of space targets are mostly non-cooperative targets, and the orbital maneuvers performed by these targets, in most cases, the relevant information is not public, but knowing these space event information is crucial for taking appropriate response measures in a timely manner, so it is of great significance to track space maneuvering targets.
[0003] To address the problem of maneuvering target tracking, various real-time maneuvering target tracking methods have emerged. These methods can be roughly divided into three types: adaptive Kalman filtering methods based on a single model, multi-model methods, and decision-based adaptive filtering methods. Decision-based adaptive filtering methods mainly include equivalent noise methods, input estimation methods, and model switching methods. Among them, the typical representative of the model switching method is the variable state dimension method (VSD), which was proposed by Bar-Shalom and Brimiwal in 1982. The algorithm uses a switch switching model to switch between two tracking models by detecting whether the target is maneuvering. The algorithm has a simple structure and low computational complexity, can provide an estimate of maneuvering acceleration, and is suitable for tracking situations with high real-time requirements.
[0004] Since space maneuvering target tracking often requires real-time performance and unknown maneuvering acceleration estimation, the VSD method has great application advantages in the field of space maneuvering target tracking. At present, many studies have proposed different space maneuvering target tracking algorithms based on variable state dimensions, verifying the feasibility and effectiveness of such space maneuvering target tracking methods.
[0005] In 2005, Richard S. Hujsak proposed a spatial non-cooperative maneuvering target tracking algorithm that adopted the idea of variable dimension. The algorithm uses the 6-dimensional target orbit state, solar light pressure coefficient and measurement deviation to form a common state variable. On this basis, the unknown maneuvering acceleration of the target is introduced as a dimension-increasing state variable, and the ratio of the residual to its covariance is used to determine the occurrence of maneuvering. The algorithm estimates the maneuvering acceleration component of the target, but in order to obtain the appropriate maneuvering acceleration uncertainty, it is necessary to use the measurement data of the target maneuvering for a period of time to repeatedly filter and iteratively search for the speed process noise amplitude. This is actually a post-processing process, which is not suitable for occasions with real-time requirements, and is only suitable for tracking targets that perform small maneuvers for a long time. In addition, Richard S. Hujsak has not yet introduced the measurement means used, so it is impossible to determine the observation occasions to which the algorithm is applicable.
[0006] In 2008, Xuanyu Hu et al., inspired by the variable dimension algorithm, independently proposed a nonlinear VSD algorithm for tracking maneuvering spacecraft. This algorithm extended the VSD algorithm based on the linear model hypothesis proposed by Bar-Shalom to the field of non-cooperative spacecraft tracking dominated by orbital dynamics, and adopted the Unscented Kalman Filter (UKF). However, this algorithm assumes that the unknown maneuvering acceleration of the spacecraft is in the same direction as its velocity, and takes the position and velocity of the target as the observed quantity, without considering the actual observation scenario, so the scope of application is relatively limited.
[0007] In 2015, Gary M. Goff and Johnathan T. Black introduced IMM in VSD filter and proposed an algorithm to track spacecraft performing continuous thrust using ground radar. The method compared the effects of two maneuver start detection methods, the residual-based test statistic and the filter smoothing consistency test, and introduced IMM in the dimensionality increase filtering process to make the filter switch between several different process noise levels to adapt to different levels of maneuver thrust. For maneuver termination detection, the maneuver termination detection effect under different models and test methods was studied according to different classification methods such as single dimensionality increase model and IMM dimensionality increase model and test methods such as residual-based test statistic, test statistic based on maneuver acceleration estimation and filter smoothing consistency test. The algorithm pointed out that the VSD-IMM tracking method can flexibly deal with unknown maneuvers of different sizes, but if the parameters of the single model VSD filter are set appropriately, its effect can be better than the multi-model method.
[0008] In 2019, Zhai Guang et al. considered the close-range relative motion scenario in circular orbit, introduced the CW dynamic equation in the VSD filter, and used space-based sensors to track close-range space targets performing unknown maneuvers. This study assumes that the observed quantity is the position and velocity of the target relative to the space-based observation platform, does not consider the actual observation means, and requires that the maneuver must be a constant, so the application conditions of this algorithm are also more.
[0009] In 2023, Zhixun Zhang et al. combined the variable dimension algorithm with IMM and proposed an algorithm suitable for tracking non-cooperative maneuvering targets in space under space-based binocular observation conditions. This algorithm is different from the VSD-IMM algorithm proposed by Gary M.Goff, which only uses IMM after successful maneuver detection. This algorithm constructs a simple extended Kalman filter (EKF) model with target position and velocity as state variables and a complex strong tracking extended Kalman filter (STEKF) model with unknown maneuvering acceleration as the increased dimensional state variable, and continuously uses the two models to interact during the tracking process. This algorithm proposes an adaptive Markov transfer matrix algorithm to realize the interaction between the two models, and the model probability determines whether the target is maneuvering. This algorithm is superior to the traditional IMM algorithm and can be used in scenarios where two cooperative spacecraft jointly track non-cooperative targets in space by optical angle observation.
[0010] In summary, most of the current research on space maneuvering target tracking algorithms based on the VSD method uses radar tracking or binocular imaging tracking. However, radar is difficult to use on space-based platforms due to its large size and high power, and binocular imaging requires at least two cooperative observation platforms. These characteristics lead to more restrictions that must be met when tracking space targets. With the continuous increase in the number of space events and the continuous diversification of their types, tracking space maneuvering targets as conveniently and accurately as possible is the key development direction of this type of research in the future. In response to this trend, the present invention proposes a real-time and accurate space maneuvering target tracking algorithm suitable for monocular observation conditions based on variable state dimension. Summary of the invention
[0011] (I) Purpose of the invention
[0012] In order to solve the problem of tracking maneuvering targets on orbit in space, the present invention proposes a monocular tracking algorithm for maneuvering targets in space based on variable state dimension. The present invention has the advantages of real-time and accurate tracking of maneuvering targets in space, can solve the monocular tracking problem under the condition of only a single optical observation platform, has few constraints on observation equipment, is highly convenient, and has a good ability to track non-cooperative maneuvering targets in space.
[0013] (II) Technical solution
[0014] Here, taking a single observation platform tracking a non-cooperative target in space through optical angle observation and performing maneuver detection on its unknown maneuvers as an example, the main technologies of the monocular tracking algorithm for space maneuvering targets based on variable state dimension described in the present invention are described in detail.
[0015] 1. Precise orbit determination before space target maneuvers
[0016] When the space (non-cooperative) target is not maneuvering, the common state variable x consisting of its position and velocity in the ECI coordinate system is used, that is,
[0017] x=[r T v T ]=[xyzv x v y v z ] T (1)
[0018] In the space target tracking algorithm, we can only consider the influence of the main perturbation force (such as the non-spherical gravity of the earth), ignore the influence of the secondary perturbation force (such as the earth tide and the gravity of the sun and the moon), and regard their joint effect as a random quantity, so that the following pseudo-random reduced dynamic model can be obtained, that is, the state variable satisfies the following continuous nonlinear time-varying state equation with random input terms:
[0019]
[0020] Among them, a m =a m (r, v, p, t) is the acceleration modeled. In addition to being related to the position of the target, it may also be related to speed, time, etc. according to the specific orbital dynamics model used. p is a dynamics model parameter, such as the atmospheric drag coefficient C D 、Solar pressure coefficient C R ;w t Represents the unmodeled acceleration, which is generally represented by zero-mean Gaussian white noise, which is the random input term.
[0021] The observed data is usually discrete, so it is necessary to discretize equation (2). The functional relationship between the state variables between two different discrete moments can be described by the following discrete nonlinear time-varying state equation:
[0022] x k =f(x k-1 ,t k ,t k-1 )+w k-1 (3)
[0023] Among them, x k Indicates t k The state variable of the spatial target at time w k-1 is the discretized process noise, and its properties are similar to those of the continuous process noise w i The nature and time interval t k -t k-1 Although f(·) exists objectively, it is generally not expressed explicitly. However, we can use equation (2) to express x by numerical integration without considering random input terms. k-1 Get x k .
[0024] In order to use the extended Kalman filter to estimate the target state through observations, the state error transfer matrix and process noise matrix are also required.
[0025] x k x k-1 The partial derivative matrix of is the orbital state transfer matrix, denoted as Φ(t k ,t k-1 ), which satisfies the following matrix differential equation:
[0026]
[0027] Its initial condition is
[0028] Φ(t k-1 ,t k-1 )=I 6×6 (5) By numerically integrating equation (4) until t k At this moment, we can get Φ(t k ,t k-1 ).
[0029] The process noise w is a discrete form k-1 The covariance matrix of is the process noise matrix Q(t k ,t k-1 ), which satisfies the following matrix differential equation:
[0030]
[0031] The third term on the right side of the equation is the continuous form of process noise at t k-1 The covariance matrix at time, σ wt w t In t k-1 The standard deviation at the moment. The initial value condition of the process noise matrix is
[0032] Q(t k-1 ,t k-1 )=0 6×6(7) By numerically integrating equation (6) until t k At this moment, we can get Q(t k ,t k-1 ).
[0033] In order to highlight the key points, the optical observation model here adopts a relatively simple direction-finding model, and takes the right ascension α and declination δ data of the unit direction vector of the target satellite relative to the observation platform in the ECI coordinate system as the observation quantities.
[0034] Assume that the position vectors of the observation platform and the target satellite in the ECI coordinate system are r O 、r T , let ρ = r T -r O , then the unit direction vector of the target satellite relative to the observation platform is
[0035]
[0036] The observed quantity is
[0037]
[0038] The measurement equation is
[0039] z=h(x)+υ (10)
[0040] Where h(x) is a nonlinear function of the state variable x, υ is the measurement noise, and its covariance matrix is R. It is easy to know that the relationship between the observed quantity and the relative unit direction vector satisfies
[0041]
[0042] In the filtering process, we need to use Taylor expansion to linearize the measurement equation, and the Jacobian matrix of the measurement equation is:
[0043]
[0044] Based on formula (11), we can further get
[0045]
[0046] Assume that the state estimate after measurement update at time k is The state estimation error covariance matrix is According to equations (4) to (7), we can perform numerical integration to obtain the state transfer matrix Φ k =Φ(t k+1 ,t k ) and the process noise matrix Q k =Q(t k+1 ,t k), so the one-step prediction equation at time k+1 is
[0047]
[0048] Get the observed value z k+1 After that, the measurement update can be performed. According to formula (12), the Jacobian matrix of the measurement equation at this moment can be calculated as:
[0049]
[0050] Therefore, the measurement update equation is
[0051]
[0052] In this way, the state estimation of the spatial target and its error covariance at time k+1 can be obtained.
[0053] 2. Real-time mobile detection
[0054] By obtaining the state estimate after the measurement update at time k+1, we can obtain the observation prediction
[0055]
[0056] In order to use the variable state dimension method for maneuver detection, we first introduce innovation and fading memory averaging. The difference between the observed value of a feature of an object at a certain moment and its estimated value is the measurement residual or innovation. In the scenario introduced here,
[0057]
[0058] The new information is a zero-mean white noise sequence, which is assumed to obey a Gaussian distribution. Let the covariance matrix of the new information ν(k) be S(k), and define the random variable
[0059]
[0060] It is easy to prove that ∈ v Obey z Dimensional χ 2 Distribution, where n z is the dimension of the observation vector. Then, we construct the recursive sequence of random variables
[0061]
[0062] Let the initial condition be And 0<α<1, we can recursively get
[0063]
[0064] The random variable in the above formula is is the fading memory sum of the new information sequence ν(k). It is a weighted statistic that integrates the new information at k moments. The farther the new information is from the current moment, the smaller the impact or weight is, and this effect decreases exponentially. Now, for the random variable After normalization, we can get
[0065]
[0066] The random variable z(k) in the above formula is the fading memory average or exponentially discounted average of the new information sequence ν(k). It is easy to see that
[0067]
[0068] According to the above analysis of the characteristics and meaning of the fading memory and, it can be seen that it can be used as a test statistic for maneuver detection. The effective window length of the maneuver to be tested is
[0069]
[0070] if only Exceeding a certain threshold γ s , we believe that maneuvering has occurred, and the estimator switches from the static model to the maneuvering model. Otherwise, it is considered that the target has not maneuvered.
[0071] 3. State estimation based on the increased-dimensional model
[0072] Combined with formula (17), if the maneuver detection result at time k+1 determines that the target has not maneuvered, then continue the recursive calculation according to the extended Kalman filter algorithm described in step 1; if the maneuver detection result at time k+1 determines that the target has maneuvered, then return to the starting point of the window, that is, time k+1-s, and use the enhanced dynamic model after dimensionality increase to filter again. At this time, proceed as follows:
[0073] The target maneuvering acceleration a t As an additional state variable, let the component of the maneuvering acceleration in the reference coordinate system be
[0074] a t =[a tx a ty a tz ] T (25) The following increased-dimensional state variables can be obtained
[0075]
[0076] When establishing the state equation of the increased-dimensional state variable, the maneuvering acceleration is modeled as a constant value affected by random factors, that is,
[0077]
[0078] Among them, the time derivative of the common state variable considering the influence of maneuvering acceleration is
[0079]
[0080] Similarly, due to the discreteness of the measured data, equation (27) needs to be discretized. At this time, the functional relationship between the state variables between two different discrete moments needs to be described using the following discrete nonlinear time-varying state equation:
[0081]
[0082] Among them, y k Indicates t k The dimension-increasing state variable of the spatial target at time v k-1 is the process noise of the discretized enhanced state equation. Similarly, its properties are similar to the continuous form of process noise v t The nature and time interval t k -t k-1 And g(·) generally still has no explicit expression, but we can use equation (29) to numerically integrate x without considering noise. k-1 Get x k .
[0083] Similarly, in order to estimate the increased-dimensional state variables using the extended Kalman filter, the state transfer matrix and process noise matrix of the increased-dimensional state variables need to be given.
[0084] y k For k-1 The partial derivative matrix of is the orbital state transfer matrix, denoted as F(t k ,t k-1 ), which satisfies the following matrix differential equation:
[0085]
[0086] Its initial condition is
[0087] F(t k-1 ,t k-1 )=I 9×9 (31)
[0088]
[0089] From equations (30) and (31), we know that F1≡0, F2≡I 3×3 ,as well as
[0090]
[0091] From the first equation of equation (33), it can be seen that the matrix differential equation satisfied by the state transfer matrix of the ordinary state variable part of the increased-dimensional state variable is the same as equation (4) and equation (5), that is, it is consistent with the matrix differential equation satisfied by the state transfer matrix when the ordinary state variable is used in the non-maneuvering case. By numerically integrating equation (33) until t k At this moment, we can get Φ(t k ,t k-1 ) and S(t k ,t k-1 ), so we can get F(t k ,t k-1 ).
[0092] Discrete form of process noise v k-1 The covariance matrix of is the process noise matrix V(t k ,t k-1 ), which satisfies the following matrix differential equation:
[0093]
[0094] The third term on the right side of the equation is the continuous form of process noise at t k-1 The covariance matrix at time, σ νt ν t In t k-1 The standard deviation at the moment. The initial value condition of the process noise matrix is
[0095] V(t k-1 ,t k-1 )=0 9×9 (35)
[0096] remember
[0097]
[0098] Among them, the process noise covariance matrix corresponding to the maneuvering acceleration component satisfies the following matrix differential equation:
[0099]
[0100] If the time interval t k -t k-1 Always fixed to T and If nothing changes, then will always be the same constant. By numerically integrating equation (34) until t k At this moment, we can get Q(t k ,t k-1 ).
[0101] In addition, a simpler process noise model can also be used, that is, directly discretizing the noise ν in equation (29): k-1 Assuming that only the maneuvering acceleration component has random input items, the process noise matrix is
[0102]
[0103] Next, based on the optical direction finding model using ordinary state variables introduced in step 2, the measurement matrix using increased-dimensional state variables can be obtained: At this time, the measurement equation is
[0104]
[0105] From the definition of the measurement matrix, it is easy to see that
[0106]
[0107] in,
[0108]
[0109] Assume that m=k+1-s+l in the re-filtering process and the updated dimensional state is estimated as The state estimation error covariance matrix is According to equations (30) to (38) or part of them, numerical integration or other calculations can be performed to obtain the state transfer matrix F m =F(t m+1 ,t m ) and the process noise matrix V m =V(t m+1 ,t m ), so the one-step prediction equation at time m+1 is
[0110]
[0111] Use the corresponding observation z at time m+1 m+1 , the measurement can be updated. According to formula (40), the Jacobian matrix of the measurement equation at this moment can be calculated as
[0112]
[0113] Therefore, the measurement update equation is
[0114]
[0115] In this way, the state estimation of the spatial target and its error covariance at time m+1 can be obtained.
[0116] In order to enhance the algorithm's ability to cope with different levels of unknown maneuvering acceleration, when a maneuver is detected and returned to the window starting point, that is, m = k + 1-s, for re-filtering, a shorter maneuvering acceleration estimation window is selected from this moment as the starting point for the initial estimation of the unknown maneuvering acceleration. When a good initial value of maneuvering acceleration is provided, the effect of tracking the target will be greatly improved. Assume that the maneuvering acceleration estimation window length is q, and the new information in the window obtained by filtering using the dimensionality-increasing state model is
[0117]
[0118] The sum of the squares of the innovation norms within the window is used as the objective function, that is,
[0119]
[0120] The initial estimated value of maneuvering acceleration at time m = k + 1-s is used as the optimization variable to obtain the least squares optimization problem. The initial estimated value of maneuvering acceleration that minimizes the objective function J is obtained through iterative solution. This estimated value is then used as the initial estimated value of the maneuvering acceleration of the dimension-increasing model at the starting point of the window to track the next target maneuvering phase.
[0121] 4. Maneuver termination detection
[0122] Different from the conventional maneuvering termination detection strategy in the VSD filter, in order to solve the delay problem that the maneuvering start detection in the VSD filter lags behind the actual maneuvering termination moment, the maneuvering termination detection in the present invention starts from the next moment after the maneuvering starts, i.e., moment m+1.
[0123] The test statistic used to detect the occurrence of maneuvers However, this test statistic cannot be used to determine whether the maneuver model should be converted to the ordinary model, that is, it cannot be used to test whether the maneuver disappears, because when the maneuver model is used to track a target in a non-maneuvering state, the innovation may not increase. On the contrary, due to the higher accuracy, the innovation may decrease. Another possibility is that when the acceleration estimate is not accurate enough, the estimate of the ordinary state variable will deviate further from the true value, resulting in filter divergence. At this time, the innovation will continue to be high, and it is impossible to determine when the maneuver disappears.
[0124] At this point, the scheme for switching back from the maneuver model to the normal model is as follows: compare the estimated accelerations with their standard deviations. If they are not statistically significant, the maneuver hypothesis is rejected and the maneuver is considered to have disappeared. Therefore, it is also necessary to construct a test statistic for the acceleration estimate, namely
[0125]
[0126] in, is the state estimator The acceleration component in is the submatrix of the state estimation error covariance matrix composed of the acceleration error covariance. Select a window length p and construct the test statistic
[0127]
[0128] The above statistic is a moving average over a window length of p. This statistic combines the estimated acceleration values at each moment within the window length p. Therefore, this statistic can be used to determine whether the acceleration is close to 0, that is, whether the maneuver disappears. The test statistic here can also use the fading memory average. When the statistic is less than a certain threshold γ a When the acceleration is considered to be insignificant, the maneuver is determined to have disappeared, and the normal dynamics model is switched back to continue filtering using normal state variables. If the maneuver is not determined to have terminated, the enhanced dynamics model after dimensional increase is continued to be used for filtering.
[0129] (III) Advantages
[0130] The advantages of the monocular tracking algorithm for space on-orbit maneuvering targets based on variable state dimension provided by the present invention are:
[0131] ① The algorithm proposed in this invention has the ability to track space maneuvering targets in real time and accurately.
[0132] During the process of continuous precise orbit determination, when the target undergoes unknown maneuvers, the algorithm can obtain an initial estimate of the maneuvering acceleration, thereby continuously obtaining high-precision target state estimates, which has a very good application effect.
[0133] ② The algorithm proposed in this paper is aimed at different types of maneuvering modes that may exist in space maneuvering targets.
[0134] It has the ability to cope, so it has strong universality and adaptability.
[0135] ③ The algorithm proposed in this invention is suitable for observing space maneuvering targets under single-line-of-sight observation conditions of optical observation platforms.
[0136] Tracking meets the convenience requirements of the observation platform and expands the range of observation platform types with tracking capabilities. BRIEF DESCRIPTION OF THE DRAWINGS
[0137] Figure 1 It is a flowchart of the implementation steps of the present invention
[0138] Figure 2 This is the flow chart for switching between normal and increased-dimensional models DETAILED DESCRIPTION
[0139] The following will be combined with the attached Figure 1 and technical solutions, taking a single observation platform tracking a non-cooperative target in space through optical observation and performing maneuver detection on its unknown maneuvers as an example, the specific implementation process of the present invention is further described in detail.
[0140] Step 1: Continuously perform precise orbit determination and real-time maneuver detection on space targets whose maneuvers are not detected
[0141] The position and velocity in the ECI coordinate system are used as the common state variable x of the space (non-cooperative) target when it is non-maneuverable, that is,
[0142] x=[r T v T ]=[xyzv x v y v z ] T (49)
[0143] It satisfies the following continuous nonlinear time-varying state equation with random input terms:
[0144]
[0145] For the following matrix differential equation in the time interval [t k ,t k+1 ], we can obtain the state transfer matrix Φ from the kth to the k+1th observation time. k =Φ(t k+1 ,t k ):
[0146]
[0147] For the following matrix differential equation in the time interval [t k ,t k+1 ], we can obtain the process noise matrix Q from the kth to the k+1th observation time. k =Q(t k+1 ,t k ):
[0148]
[0149] The Jacobian matrix of the measurement equation is calculated as follows:
[0150]
[0151] in,
[0152]
[0153] Assume that the state estimate after measurement update at time k is The state estimation error covariance matrix is The one-step prediction result at time k+1 is calculated as follows:
[0154]
[0155] The measurement update result is further calculated as follows:
[0156]
[0157] Then calculate the new interest as follows:
[0158]
[0159] The observed quantity prediction
[0160] Then calculate the following statistics:
[0161]
[0162] Depend on Recursively, we can get the decay memory sum of the new information sequence at time k+1:
[0163]
[0164] Set a certain threshold γ s ,if only If the threshold is exceeded, we consider that a maneuver has occurred, and we return to the starting point of the window and re-filter using the enhanced dynamics model after dimensionality increase; otherwise, we consider that the target has not maneuvered, and continue to recursively calculate using the extended Kalman filter algorithm according to the ordinary model.
[0165] In addition, if the current filtering has reached the tracking stop condition, such as reaching the maximum tracking time, the tracking is terminated.
[0166] This process corresponds to the Figure 1 Block diagram①.
[0167] Step 2: Construct an increased-dimensional state model for the detected maneuvering target
[0168] If the maneuver detection result at time k+1 determines that the target has maneuvered, then the increased-dimensional state model should be constructed. The increased-dimensional state variable is
[0169]
[0170] The state equation of the increased-dimensional state variable is
[0171]
[0172] Among them, the time derivative of the common state variable considering the influence of maneuvering acceleration is
[0173]
[0174] This process corresponds to the Figure 1 Block diagram②.
[0175] Step 3: Orbit recovery after maneuver detection
[0176] After detecting the maneuver, it is necessary to return to the starting point of the window, that is, time k+1-s, and re-filter using the enhanced dynamics model after dimensionality increase. The window length is calculated as follows:
[0177]
[0178] For the following matrix differential equation in the time interval [t k ,t k+1 ], we can obtain the state transfer matrix F(t k ,t k-1 ):
[0179]
[0180] For the following matrix differential equation in the time interval [t k ,t k+1 By numerically integrating the process noise matrix V(t k ,t k-1 ):
[0181]
[0182] Or use the simpler process noise matrix as follows:
[0183]
[0184] The Jacobi matrix of the measurement equation of the increased-dimensional model is calculated as follows:
[0185]
[0186] The calculation method of H is the same as step 1.
[0187]
[0188] Assume that the dimension-increased state after the measurement update at time m=k+1-s+l in the re-filtering process is estimated to be The state estimation error covariance matrix is Then the one-step prediction value of the dimension-increased state at time m+1 is:
[0189]
[0190] The measurement update value can be further calculated according to the following formula:
[0191]
[0192] When the maneuver is detected and the filter is re-filtered back to the window starting point, that is, time m = k + 1-s, a shorter maneuver acceleration estimation window is selected with its length q as the starting point. The new information in the window obtained by filtering using the dimensionality-increasing state model is
[0193]
[0194] The sum of the squares of the innovation norms within the window is used as the objective function, that is,
[0195]
[0196] The initial estimated value of maneuvering acceleration at time m = k + 1-s is used as the optimization variable to obtain the least squares optimization problem. The initial estimated value of maneuvering acceleration that minimizes the objective function J is obtained through iterative solution. This estimated value is then used as the initial estimated value of the maneuvering acceleration at the start point of the window in the dimension-increasing model to filter the next target maneuvering phase.
[0197] When l=s, ie, m=k+1, the re-filtering process within the window is completed.
[0198] At the same time, during the re-filtering process, it is also necessary to calculate the value of the test statistic constructed by the acceleration estimate at each moment as follows:
[0199]
[0200] Select a window length p and calculate the value of the moving average statistic as follows:
[0201]
[0202] Set a specific threshold γa , when the moving average statistic is less than γ a When , we consider the acceleration to be insignificant and determine that the maneuver disappears, we switch back to the ordinary dynamics model and continue filtering using ordinary state variables, that is, return to step 1; if the maneuver is not determined to be terminated, we continue to use the enhanced dynamics model after dimensionality increase for filtering.
[0203] At the same time, if the current filtering has reached the tracking stop condition, such as reaching the maximum tracking time, the tracking is terminated.
[0204] This process corresponds to the Figure 1 Block diagram③.
[0205] Step 4: Real-time orbit determination after maneuver detection
[0206] When l = s, that is, m = k + 1, if the maneuver is still not determined to be terminated according to the value of the moving average statistic, then with the arrival of the next observation moment, the filtering process will enter the real-time precise orbit determination stage of the maneuvering state. In this stage, the observation data used is the real-time data received from the new moment, and the specific calculation process is consistent with step 3. At the same time, in this process, it is still necessary to calculate the moving average statistic at each observation moment in the real-time orbit determination process to determine whether the maneuver disappears, so as to determine whether to switch back to the normal model, and it is necessary to determine whether the current filtering has reached the tracking stop condition. If so, the tracking should be ended.
[0207] This process corresponds to the Figure 1 Block diagram④.
Claims
1. A monocular tracking algorithm for space on-orbit maneuvering targets based on variable state dimension, taking the tracking of space non-cooperative maneuvering targets by a single observation platform through optical single line of sight as an example, characterized by: Step 1: Continuously perform precise orbit determination for space targets with no detected maneuvers When the space (non-cooperative) target is non-maneuverable, the common state variables and orbital dynamics model consisting of its position and velocity in the ECI coordinate system are used, and the right ascension and declination of the target's relative line of sight vector relative to a single observation platform are used as observation quantities, and real-time precise orbit determination is performed according to a certain filtering method such as the extended Kalman filter. Step 2: Real-time maneuver detection during precise orbit determination of space targets The recursive random variable sequence based on the new information is the fading memory sum of the new information sequence. In the process of precise orbit determination of space targets without detected maneuvers, the fading memory sum is calculated at every moment. After setting a certain detection threshold, the value of the fading memory sum is compared with the detection threshold to determine whether the maneuver occurs. Step 3: Use the dimensionality-increasing model to estimate the state of the detected maneuvering target If the maneuver detection result at a certain moment determines that the target has not maneuvered, the extended Kalman filter algorithm described in step 1 continues to recursively calculate; if the maneuver detection result at this moment determines that the target has maneuvered, it should return to the starting point of the window and re-filter using the enhanced dynamic model after dimensionality increase. In order to enhance the algorithm's ability to cope with different levels of unknown maneuvering acceleration, when a maneuver is detected and re-filtered by returning to the starting point of the window, a shorter maneuvering acceleration estimation window is selected with this moment as the starting point, and the initial estimated value of the maneuvering acceleration at the starting point is used as the optimization variable, and the sum of the squares of the innovation norm in the window is used as the objective function. The initial estimate of the unknown maneuvering acceleration is obtained through optimization. Among them, the increased-dimensional state model uses the maneuvering acceleration of the target as an additional state variable to expand the dimension of the ordinary state variable. By modeling the maneuvering acceleration state equation, the state equation model of the increased-dimensional state variable can be obtained. Step 4: Perform maneuver termination detection on the detected maneuvering target In the process of detecting maneuvers and filtering with the increased-dimensional model, maneuver termination detection is performed at every moment. First, a test statistic for maneuver termination determination needs to be constructed, such as a moving average test statistic constructed based on acceleration estimation. Secondly, the value of the test statistic for maneuver termination determination needs to be calculated based on the observed values or state estimates at each moment, and the statistic is compared with the set threshold to determine whether the maneuver is terminated. If the maneuver is terminated, switch back to the normal model and execute according to step one, otherwise continue to use the increased-dimensional model and execute according to step three.
