Space target orbit transaction behavior intelligent identification method based on satellite-borne passive measurement

Abstract

The invention discloses a space target orbit transaction behavior intelligent identification method based on satellite-borne passive measurement, and the method comprises the steps: carrying out the data preprocessing through the feature extraction capability of multi-dimensional singular spectrum analysis, putting forward a two-dimensional fusion feature quantity sequence as an input value, taking the maneuvering state quantity of a target satellite as an output value, and carrying out the recognition of the target orbit transaction behavior. And training the long-short-term memory neural network offline, establishing a maneuvering detection dichotomy model and deploying the maneuvering detection dichotomy model on a satellite for online use, thereby realizing maneuvering detection of any target on a GEO type orbit. The space-based non-cooperative target maneuvering detection method does not need other prior information and hypothesis conditions, can perform maneuvering detection judgment on the space-based non-cooperative target only by using the two-dimensional angle measurement sequence output by the space-based angle measurement system, takes the periodic feature component extracted by MSSA preprocessing as the input of the LSTM, eliminates high-frequency noise through orthogonalization decomposition, and improves the detection accuracy of the space-based non-cooperative target. The LSTM focuses on maneuver-related time sequence mode learning, and the problem of misjudgment caused by frequency band overlapping of noise and maneuver signals in a traditional method is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of space-based passive detection, and particularly to an intelligent recognition method for the orbit abnormal movement behavior of space targets in spaceborne passive measurement. Background Art

[0002] With the increasing number of satellites and non-cooperative targets and the increasingly fierce competition in space militarization, the safe and stable operation of on-orbit spacecraft has been severely challenged. The space resources of communication and navigation satellites are particularly tense, and the resulting space security issues are intensifying. Space security means protecting space systems from interference and damage caused by natural or human factors. At present, the number of various spacecraft is increasing exponentially, the key orbital regions are becoming more and more crowded, satellites of different countries and for different purposes are shuttling in space, and space debris is also increasing. These non-cooperative targets have seriously affected the on-orbit safe operation of spacecraft, and more and more problems are induced by various potential unsafe factors. How to enhance the capabilities of one's own spacecraft in aspects such as perception, monitoring, and decision-making has become an urgent problem to be solved in the field of space situation awareness. In this context, for non-cooperative targets, it is particularly important to have the ability to detect, perceive, and take actions on their orbital pulse maneuvers in a timely manner.

[0003] At present, the observation systems for space situation awareness are becoming increasingly diverse. According to the installation locations of the observation facilities, they can be mainly divided into the following three categories: space-based, ground-based, and sea-based observation systems. Among them, for the detection task of space-based non-cooperative targets, the relative orbit determination system based on passive angle-only measurement by an optical camera is widely used. Compared with other commonly used spaceborne measurement sensors, an optical camera has the characteristics of high concealment, low energy consumption, simplicity, convenience, and reliability, and is particularly suitable for completing the passive detection task of space non-cooperative targets.

[0004] In the initial orbit determination stage of the passive measurement of space non-cooperative targets, it is very likely that the initial orbit determination error becomes larger or even fails due to the orbital maneuver of the target. Therefore, to achieve high-precision initial orbit determination, it is necessary to detect the orbital maneuver behavior of the target in the initial orbit determination measurement stage. However, in the initial orbit determination stage, there is a lack of prior information on the orbital motion of the target, and the only available observation data is the line-of-sight angle information of the target, which makes the maneuver detection task in this stage very difficult. How to realize the target orbital maneuver detection starting from the analysis of the original line-of-sight angle measurement sequence is a problem that must be solved.

[0005] Existing time series anomaly detection methods for passive angle measurement include Discrete Wavelet Transform (DWT), Principal Component Analysis, Empirical Mode Decomposition (EMD), Isolation Forest, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), etc. Although these existing algorithms have the ability to detect abnormal maneuver data sequences of targets, they require relatively complete prior information data or need to perform calculations under specified hypothesis testing models or special orbits. Therefore, in the scenario of maneuver detection for non-cooperative targets, it is difficult to apply the detection model based on a single algorithm. Summary of the Invention

[0006] Aiming at the problems existing in the existing anomaly detection algorithms, the present invention provides an intelligent recognition method for the orbital anomaly behavior of space targets in spaceborne passive measurement. In the case of insufficient prior information, only angle data is used to continuously and accurately detect and discriminate the maneuvers of targets in real time under the condition of lightweight spaceborne computing load.

[0007] An intelligent recognition method for the orbital anomaly behavior of space targets in spaceborne passive measurement includes the following steps:

[0008] Step 1: Install an optical camera at the centroid of the sensing satellite, establish a relative line-of-sight measurement model of the optical camera in the local vertical local horizontal coordinate system, and obtain a two-dimensional relative line-of-sight measurement angle sequence as the detection quantity sequence;

[0009] Step 2: Define a data preprocessing algorithm based on multi-channel singular spectrum analysis, define appropriate safety intervals and sliding window length values, and perform multi-channel singular value decomposition on the two-dimensional detection quantity sequence within the current sliding window, remove trend components, irregular components, etc., and calculate the reconstructed periodic feature components;

[0010] Step 3: Define a maneuver detection method based on a long short-term memory neural network, determine appropriate parameters, and offline train the neural network with the standardized data to obtain a binary classification model based on the feature sequence;

[0011] Step 4: Deploy the angle-only intelligent maneuver detection algorithm on the sensing satellite, input the relative measurement angle measured by the optical camera into the algorithm, and realize the online discrimination of the maneuver detection of the target satellite.

[0012] Preferably, the relative line-of-sight measurement model of the optical camera established in Step 1 is specifically: , where respectively represent the position vectors of the tracking spacecraft A and the target spacecraft B in the LVLH system at a certain moment.

[0013] Preferably, the P-dimensional angle measurement time series obtained in step 1 is , where N is the length of the time series.

[0014] Preferably, the multi-channel singular spectrum analysis algorithm in step 2 does not require prior conditions such as assuming parameter models and assuming stationarity conditions. It is applicable to the initial orbit determination stage with insufficient prior information. By decomposing and reconstructing the trajectory matrix of the time series to be studied, different components in the time series are extracted, thus completing tasks such as feature extraction and data preprocessing.

[0015] Preferably, step 2 is specifically as follows: Step 2.1, after setting the length of the safety interval, perform sliding window processing and time delay sorting on the measurement angle sequence obtained in step 1 to obtain a trajectory matrix; Step 2.2, perform multi-channel singular value decomposition on the trajectory matrix and arrange it in descending order of singular values to obtain the time empirical orthogonal functions corresponding to different trend components; Step 2.3, calculate the correlation coefficient to remove trend components, irregular components, etc. and calculate the reconstructed characteristic components; Step 2.4, perform normalization processing on the characteristic components and then perform standardization processing.

[0016] Preferably, in step 2.1, the dimension of the angle measurement sequence is P, the sliding window length is set to N, the number of rows of the trajectory matrix is selected as L, and the two-dimensional detection sequence is arranged with a time lag. Let , then the trajectory matrix is a matrix of: , which is composed of hysteresis vectors of length : Therefore, the combined multi-channel trajectory matrix is obtained.

[0017] Preferably, in step 2.2, to avoid problems such as different dimensions and large computational amounts that may be encountered in singular value decomposition, an equivalent substitution method is selected to calculate the square matrix of the trajectory matrix: , perform eigenvalue decomposition on it to obtain eigenvalues: and the corresponding eigenvectors: , reflects the evolution type of the time series and is called the time empirical orthogonal function.

[0018] Preferably, in step 2.3, the time empirical orthogonal functions of different trend components obtained by decomposition are calculated pairwise for the correlation coefficient. For the time series , its weighted correlation coefficient is expressed as: . Calculate the cross-correlation coefficients of L temporal empirical orthogonal functions and visualize them - draw the heatmap of the weighted correlation coefficients.

[0019] Preferably, the heatmap of the weighted correlation coefficients is only used as part of the visual analysis. The specific selection process is as follows: First, set the cross-correlation coefficient threshold η = 0.83. When the cross-correlation coefficient satisfies , the two sequences are regarded as a group, arranged in descending order of eigenvalues, and the corresponding subsequence group number is denoted as z. Remove the high-frequency trend components with z = 1 and the irregular components with z ≥ 4, and retain the weighted sum of the sequences within the groups of z = 2 and z = 3, corresponding to the abnormal feature components of the current sliding window sequence in turn; Second, calculate the projection of the lag vector on , that is, the temporal principal component: , and perform reconstruction through the temporal empirical orthogonal function and the temporal principal component: .

[0020] Preferably, in step 2.4, the normalization process is: , where x’ is the normalized data, x is the original data, x min , x max are the minimum and maximum values in the original data respectively; the standardization process is: , where x’’ is the standardized data, and μ and σ are the mean and standard deviation of the normalized data respectively.

[0021] Preferably, in step 3, calibrate the maneuver state quantity of the characteristic sequence: when the target does not perform an orbital maneuver, the corresponding label of the angle measurement sequence is "0", and when the target performs an orbital maneuver, the corresponding label of the angle measurement sequence is "1". The standardized characteristic components in step 2.4 are used as the input quantity of the long short-term memory neural network; the specified target maneuver state label value is used as the output quantity of the long short-term memory neural network to establish a maneuver detection binary classification model.

[0022] Preferably, the number of hidden layers of the long short-term memory neural network is 3. Among them, the long short-term memory network layer forgets or remembers the historical sequence information through the added forget gate, input gate, and output gate to complete the update of the network cell unit state. The fully connected layer saves the useful feature information and outputs it to the softmax layer. The gradient descent optimization algorithm uses the Adam algorithm, and the loss function uses MSE.

[0023] Preferably, in the binary classification problem, the activation function adopted by the softmax layer is the softmax function in the following form: , the function can normalize the output of the fully connected layer, convert the time series processing result into two probability values ranging from (0, 1) and with a sum of 1, corresponding to the probabilities that the model classifies the target as non-maneuvering ("0") and maneuvering ("1"). In addition, during the model training process, a cross-entropy loss function in the following form is selected to judge its training degree:

[0024] ,

[0025] In the formula: is the true label of the th training sample, corresponding to "0" or "1"; is the probability that the model predicts the th sample as the "1" class, that is, the probability of judging the target as maneuvering.

[0026] Preferably, the gradient descent optimization algorithm adopts the Adam algorithm, specifically: update the gradient weight ω t : , where t is the number of times, is the correction value of m t , is the correction value of v t , m t , v t are the exponential moving average of the gradient and the squared gradient respectively.

[0027] Preferably, in order to accelerate the gradient update and the stability of the algorithm, for the calculation of the gradient weights m t , v t and their correction values , , multiply by a quantity and add a constant quantity when not corrected, and then divide by this quantity when corrected, specifically: , where β1 and β2 are constants used to control the exponential decay, m t is the exponential moving average of the gradient, v t is the squared gradient, g t is the first derivative. Learn the following network specific parameters through offline training, including: learning rate decay period, learning rate decay coefficient, maximum number of iterations, number of hidden layers, initial learning rate, number of hidden layer nodes.

[0028] Preferably, in practical applications, in step 4, the normalized value of the two-dimensional feature sequence at the moment is used as the input of the long short-term memory neural network. After the time series processing of the LSTM layer, the corresponding processing result is obtained by the fully connected layer, and then classified by the softmax layer. Finally, the result label of the maneuver detection is output by the classification layer, where "0" represents that the target does not perform an orbital maneuver, and "1" represents that the target has an orbital maneuver.

[0029] Beneficial effects:

[0030] (1) In the present invention, the periodic feature components extracted by MSSA preprocessing are used as the input of LSTM. High-frequency noise is eliminated through orthogonal decomposition, enabling LSTM to focus on the learning of maneuver-related time series patterns and addressing the misjudgment problem caused by the overlapping frequency bands of noise and maneuver signals in traditional methods.

[0031] (2) The LSTM network in the present invention dynamically models the time series features after MSSA preprocessing, analyzes the long-period correlation signals caused by maneuver behaviors, and finally outputs the binary classification result of the target maneuver probability. It couples the MSSA component screening and the LSTM gate weight allocation, generating a collaborative noise reduction effect. At the same time, it can capture the micro-amplitude orbit offset caused by maneuvers, with higher detection sensitivity.

[0032] (3) The present invention solves the problem of maneuver detection in the relative orbit determination process due to the lack of prior information in space-based passive angle-only measurement data. It quantifies and analyzes the angle measurement sequence using the characteristics of multi-channel singular spectrum analysis, such as low requirements for prior information and strong data feature extraction ability, and constructs a binary classification model through the powerful time series data processing ability of the long short-term memory neural network to achieve real-time maneuver detection of the target. This maneuver detection method avoids manual threshold selection and effectively improves the success rate of maneuver detection by adopting the method of offline training and online application, with higher applicability in engineering practice.

[0033] (4) The present invention preprocesses the training data by normalizing and standardizing it, which facilitates subsequent data processing and ensures faster convergence during program operation.

[0034] (5) Through the method of multi-channel singular spectrum analysis, the present invention conducts coherence analysis on a single angle measurement sequence under different orbit types and including various pulse maneuver magnitudes, thereby eliminating the influence of noise such as camera angle measurement errors.

[0035] (6) The present invention is applicable to GEO type orbits, has a wider degree of freedom for orbit guidance, and does not require additional fuel costs.

[0036] (7) The present invention does not need to change the traditional structure of satellite sensors and is not limited by the relative distance of the satellites, with a wider degree of freedom for relative target maneuver detection. Description of the drawings

[0037] Figure 1 It is a geometric schematic diagram of space-based optical camera measurement for an embodiment of the present invention;

[0038] Figure 2 It is a logic flow chart of the maneuver detection algorithm for an embodiment of the present invention;

[0039] Figure 3 The structure diagram of the long short-term memory neural network according to an embodiment of the present invention;

[0040] Figure 4 The observation angle sequence under two conditions of maneuvering and non-maneuvering according to an embodiment of the present invention;

[0041] Figure 5 The performance index change curve of the classification model according to an embodiment of the present invention;

[0042] Figure 6 The false alarm rate and missed alarm rate change curve according to an embodiment of the present invention. Detailed implementation manners

[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0044] The present invention discloses an intelligent recognition method for the orbital abnormal behavior of space targets in spaceborne passive measurement. Aiming at the time series anomaly detection algorithms currently applied to space-based passive angle-only measurement systems, such as principal component analysis, wavelet transform, empirical mode decomposition, etc., which have high requirements for prior information and require a large amount of historical information as input to complete the accurate screening of maneuvering moments. It is common to use rich orbital parameters as test statistics in the tracking and filtering stage to achieve maneuver detection. However, in the initial orbit determination stage, a set of observables in space-based angle-only observations only has two angle values, without prior conditions such as assumed parameter models and assumed stationarity conditions. The application of the above single anomaly detection algorithm is difficult; therefore, the feature extraction ability of multi-dimensional singular spectrum analysis is used for data preprocessing to screen out effective anomaly components, and a two-dimensional fusion maneuver anomaly feature quantity sequence is proposed as the input value to improve the classification accuracy. The maneuver state quantity of the target satellite is used as the output value, and the long short-term memory neural network is trained offline to establish a maneuver detection binary classification model and deploy it on the satellite for online use, so as to realize the adaptive maneuver detection of any target on the GEO type orbit. The present invention does not require other prior information and assumed conditions, and only uses the two-dimensional angle measurement sequence output by the space-based angle measurement system to judge the maneuver detection of space-based non-cooperative targets, improving the classification accuracy and success rate.

[0045] In the initial orbit determination stage, an intelligent recognition method for the orbital abnormal behavior of space targets in spaceborne passive measurement is proposed, including the following steps:

[0046] Step 1: Install the optical camera at the centroid of the sensing satellite, establish an optical camera relative line-of-sight measurement model in the local vertical local horizontal coordinate system, and obtain a P-dimensional angle measurement time series as . In this embodiment, a two-dimensional relative line-of-sight measurement angle sequence is used as the detection quantity sequence.

[0047] The optical camera relative line-of-sight measurement model is as shown in Figure 1 , and its specific expression is: , where respectively represent the position vectors of the tracking spacecraft A and the target spacecraft B at the moment in the LVLH system.

[0048] Step 2: As shown in Figure 2 , construct a multi-channel singular spectrum analysis algorithm: By decomposing and reconstructing the trajectory matrix of the two-dimensional observation angle sequence, extract the characteristic quantity sequence in the time series. On the basis of data preprocessing, further extract the abnormal components caused by maneuvers to improve the accuracy of the classification model and reduce the classification difficulty, so as to complete tasks such as feature analysis and denoising. Define an appropriate safety interval and sliding window length value, and perform multi-channel singular value decomposition on the two-dimensional detection quantity sequence within the current sliding window to obtain the time series evolution types of different trend components, and complete feature extraction and preprocessing. The multi-channel singular spectrum analysis algorithm removes the trend components and noise components, and also screens the abnormal feature components caused by maneuvers while performing preprocessing. Specifically,

[0049] Step 2.1: Perform sliding window processing on the angle sequence obtained in Step 1. The dimension of the angle measurement sequence is P, and the sliding window length is set to N. Embed the detection sequence in each sliding window in turn, and obtain the trajectory matrix by lag arrangement: Select the number of rows of the trajectory matrix as L, perform lag arrangement on the detection sequence, and let , then the trajectory matrix is matrix: , which is composed of lag vectors of length : . In the lag matrix, is the data of one time window. Combining the obtained multi-channel trajectory matrix is expressed as: .

[0050] Step 2.2, perform singular value decomposition on the trajectory matrix, arrange the decomposition results in descending order of singular values, and obtain the temporal empirical orthogonal functions representing different trend components; among them, the orthogonal functions on each eigen-dimension correspond to specific singular values, representing an independent temporal variation pattern, and respectively resolve the periodic, trend, and noise components in the original time series. The high-weight dimensions represent significant trend components such as orbital maneuver signals, and the low-weight dimensions represent noise and minor perturbations.

[0051] To avoid the problems of different dimensions and large computational amounts that may be encountered in singular value decomposition, an equivalent alternative method is selected to calculate the square matrix of the trajectory matrix: , perform eigenvalue decomposition on it to obtain the eigenvalues: and the corresponding eigenvectors: , reflects the evolution type of the time series and is called the temporal empirical orthogonal function.

[0052] Step 2.3, calculate the correlation coefficients pairwise for the temporal empirical orthogonal functions of different trend components obtained by decomposition. For a time series, the specific calculation method of the weighted correlation coefficient is as follows: for the time series , its weighted correlation coefficient . Calculate the cross-correlation coefficients of L temporal empirical orthogonal functions (decompose the time series of the entire window into L sub-series, corresponding to L eigenvalues and eigenvectors, calculate the correlation coefficients between every two temporal empirical orthogonal functions, and draw a heat map), select the sub-series with the weighted correlation coefficient as a group, and record the group number as z. Remove the high-frequency trend components of and the irregular components of , and respectively perform weighted summation on the sequences within and groups, corresponding to the maneuver abnormal feature components of the current sliding window sequence in turn.

[0053] Select characteristic components: calculate the projection of the lag vector on , that is, the temporal principal component: , and perform reconstruction through the temporal empirical orthogonal function and the temporal principal component: .

[0054] Step 2.4, perform normalization and standardization. Among them, the normalization process is: , where is the data after normalization, is the reconstructed characteristic component, , are respectively the minimum and maximum values in the reconstructed characteristic component; the standardization process is: , where It is the standardized data, where μ and σ are the mean and standard deviation of the normalized data respectively.

[0055] Step 3: Calibrate the maneuver state quantity of the feature sequence. When the target does not perform an orbital maneuver, the angle measurement sequence corresponds to the label "0", and when the target performs an orbital maneuver, the angle measurement sequence corresponds to the label "1". The standardized characteristic components in Step 2.4 are used as the input of the long short-term memory neural network; the specified target maneuver state label value is used as the output of the long short-term memory neural network to establish a binary classification model for maneuver detection. Set the number of hidden layers of the long short-term memory neural network to 3. As Figure 3 shown, input layer: The input is a two-feature component sequence, corresponding to the periodic feature components after MSSA preprocessing (groups with z = 2 and z = 3); LSTM layer: Unfolded according to the time series to model the temporal correlation ( ). The long short-term memory network layer updates the state of the network cell unit by forgetting or remembering the historical sequence information through the added forget gate, input gate, and output gate. Dynamically adjust the weights of the historical state through the forget gate and input gate to capture the micro-offset of the orbital maneuver; fully connected layer: Map the high-dimensional features output by the LSTM to the classification space, integrate the LSTM hidden states at different time steps, and enhance the sensitivity to sudden maneuvers; the fully connected layer preserves the useful feature information and outputs it to the softmax layer to output the binary classification probability: 0: non-maneuvering, 1: maneuvering.

[0056] The gradient descent optimization algorithm uses the Adam algorithm, and the loss function uses MSE. The activation function adopted by the softmax layer in the long short-term memory neural network is the softmax function in the following form: , the function can normalize the output of the fully connected layer and convert the time series processing result into two probability values within the range and with a sum of 1, corresponding to the probabilities that the model classifies the target as non-maneuvering ("0") and maneuvering ("1"). In addition, during the model training process, select the cross-entropy loss function (Cross Entropy Loss Function) in the following form to judge its training degree: , where: is the true label of the th training sample, corresponding to "0" or "1"; is the probability that the model predicts the th sample as "1", that is, the probability of judging the target as maneuvering. The gradient descent optimization algorithm uses the Adam algorithm, specifically: update the gradient weight ω t : , where t is the number of times, is the correction value of m t , is v tCorrection value, m t , v t are the exponentially weighted moving average of the gradient and the squared gradient, respectively. To accelerate the gradient update and the stability of the algorithm, for the gradient weights m t , v t and their correction values , , during the calculation without correction, multiply by a quantity and add a constant quantity, and then divide by this quantity during correction. Specifically: , where β1 and β2 are constants used to control the exponential decay, m t is the exponentially weighted moving average of the gradient, v t is the squared gradient, and g t is the first derivative. The specific parameters of the optimal network are learned through offline training as follows: the learning rate decay period is 10, the learning rate decay coefficient is 0.1, the maximum number of iterations is 1000, the number of hidden layers is 1, the initial learning rate is 0.01, and the number of hidden layer nodes is 64.

[0057] Take the normalized value of the two-dimensional feature sequence at time as the input of the long short-term memory neural network. After the temporal processing of the LSTM layer, the corresponding processing result is obtained by the fully connected layer, and then the softmax layer is used for classification. Finally, the result label of the maneuver detection is output by the classification layer. "0" represents that the target does not perform an orbital maneuver, and "1" represents that the target has an orbital maneuver.

[0058] Deploy the angle-only maneuver detection algorithm on the sensing satellite, and input each group of relative measurement angles measured by the optical camera for a certain continuous time into the algorithm. The threshold is adaptively selected, and multi-dimensional data fusion is judged to realize the online discrimination of the maneuver detection of the target satellite.

[0059] The feasibility of the present invention will be illustrated by the following examples.

[0060] Set the following calculation conditions and technical parameters:

[0061] 1) The semi-major axis of the orbit of sensing satellite A is 42166.3 km, the eccentricity is 0.0005, the orbital inclination is 0°, the argument of perigee is 0°, the right ascension of the ascending node is 0°, and the true anomaly is 0°;

[0062] 2) The semi-major axis of the orbit of the target satellite is 42066.3 km, the eccentricity is 0.0005, the orbital inclination is 0 / 0.5°, the argument of perigee is 0°, the right ascension of the ascending node is 0°, and the true anomaly is 0°;

[0063] 3) The angle measurement accuracy of the camera is set to 0.005° ;

[0064] 4) The initial relative distances between the target satellite and the sensing satellite are 500 / 625 / 750 / 875 / 1000 km respectively.

[0065] The angle-only relative orbit determination method based on the present invention is verified by simulation with the above-set calculation conditions and technical parameters.

[0066] As Figure 4 shown is the change diagram of the angle sequence under two states of the target satellite's orbital inclination target being 0° and with and without maneuvering. When the target maneuvers, the maneuvering amount is fixed at 1 m / s, the initial relative distance of the target satellite is 1000 km respectively, and the angle measurement accuracy of the optical camera is 0.005°. As Figure 4 shown, there are large mutations in the two-dimensional angle measurement sequences at the moment of target maneuvering, and they become straight lines after the mutations.

[0067] As Figure 5 shown is the change curve diagram of the detection performance index of the classification model when the target satellite's orbital inclination target is 0°, the satellite maneuvering amount is fixed at 1 m / s and the direction is random, and the initial relative distances of the target satellite are 500 / 625 / 750 / 875 / 1000 km respectively. As Figure 6 shown is the corresponding change curve diagram of the false alarm rate and missed alarm rate indexes of the classification model.

[0068] From Figures 5 - 6 the curves in, it can be seen that by the method of this article, the success rate of detecting the maneuvering of non-cooperative targets is relatively high. Under the method of this article, the detection performance of the classification model can reach a maneuvering amount of 1 m / s magnitude and 0.005° the angle measurement accuracy of the optical camera. The overall detection success rate is above 92%, which is more than 7% higher than the accuracy of the existing methods. At the same time, both the false alarm rate and the missed alarm rate can be controlled within a small range.

[0069] Therefore, by adopting the method of the present invention, only relying on the passive angle measurement of the spaceborne optical camera can realize the maneuvering detection in the initial orbit determination stage of non-cooperative targets without prior information. Compared with the traditional single maneuvering detection method, this method can accurately detect non-cooperative targets of various relative orbit types under the conditions of lack of prior information and ranging information. At the same time, combining the data extraction ability of the multi-channel singular spectrum analysis algorithm and the time series data classification ability of the LSTM network further improves the effectiveness of the classification model. The simulation results fully demonstrate the breakthrough progress brought by the method of the present invention.

[0070] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for intelligently identifying orbital motion behavior of space targets measured by satellite-borne passive measurement, characterized in that: The following steps are involved: Step 1: construct a relative line of sight measurement model in a local vertical and local horizontal coordinate system by installing an optical camera at the centroid of the sensing satellite to obtain a P-dimensional relative line of sight measurement angle sequence; Step 2: perform trajectory matrix generation, singular value decomposition and feature reconstruction on the P-dimensional measurement angle sequence using multi-channel singular spectrum analysis MSSA to extract the denoised periodic feature components; Step 3, inputting the periodic characteristic component sequence into a long short-term memory neural network LSTM, and the LSTM learns the temporal correlation between the periodic characteristic component and the maneuvering state through offline training, and outputs a binary classification result of the target maneuvering probability to reduce noise interference and improve classification accuracy; Step 4: Deploy the trained model on the onboard computing unit to detect the target orbital maneuvering behavior in real time based on the input optical camera measurement angle sequence.

2. The method according to claim 1, characterized in that In step 1, the relative line of sight measurement model is expressed as: ,in , They represent the tracking spacecraft A and the target spacecraft B in the LVLH system. i The position coordinates at the time, represents the position vector of the tracking spacecraft, , To measure noise.

3. The method according to claim 1, characterized in that In step 2, the multi-channel singular spectrum analysis algorithm includes: step 2.1, sliding window processing and time delay sorting of the two-dimensional angle measurement sequence to obtain a trajectory matrix; step 2.2, singular value decomposition and sorting of the trajectory matrix to obtain time empirical orthogonal functions corresponding to different feature dimensions; step 2.3, calculating the weighted correlation coefficients between the time empirical orthogonal functions to remove trend components and irregular components and screen and reconstruct feature components; step 2.4, normalizing and standardizing the reconstructed feature components in turn.

4. The method according to claim 3, characterized in that In step 2.1, set the dimension of the angle measurement sequence P, the length of the sliding window N, and the number of rows in the trajectory matrix L. Perform lag arrangement, let , then the trajectory matrix Expressed as The matrix: , include The length is The hysteresis vector : ; The multi-channel trajectory matrix is ​​expressed as: .

5. The method according to claim 4, characterized in that In step 2.2, by constructing a square matrix of trajectory matrix Perform eigenvalue decomposition and obtain the eigenvalue and the corresponding eigenvector , time empirical orthogonal function , where n is the number of non-zero singular values, , which reflects the temporal variation characteristics of the time series within the sliding window.

6. The method according to claim 5, characterized in that In step 2.3, a time empirical orthogonal function represents an independent time series variation pattern. For the time series corresponding to the two time empirical orthogonal functions , weighted correlation coefficient The calculation method is: , calculate the mutual correlation coefficients of L time empirical orthogonal functions and visualize them - draw a heat map of their weighted correlation coefficients.

7. The method according to claim 6, characterized in that In step 2.3, the weighted correlation coefficient heat map selection process is as follows: set the mutual correlation coefficient threshold η, screen The sequence groups are sorted by eigenvalue size, and the corresponding subsequence group number is recorded as z. The high-frequency trend component of z=1 and the irregular component of z≥4 are removed, and the groups of z=2 and z=3 are retained as abnormal characteristic components. The weighted sum of the sequences in the group corresponds to the abnormal characteristic components of the current sliding window sequence in turn. Next, calculate the hysteresis vector exist The projection on , that is, the time principal component: , reconstructed by time empirical orthogonal functions and time principal components: , where m in the superscript and subscript represents the ordinal number of the trend component.

8. The method according to any one of claims 1 to 7, characterized in that: In step 3, the LSTM network contains 3 hidden layers, and the activation function is Softmax. , the time series processing results are converted into two probability values ​​with a range between and a sum of 1, corresponding to the model classifying the target as non-maneuvering "0" and maneuvering "1"; the loss function is cross entropy, and the optimization algorithm is Adam; whether the target performs orbital maneuvering behavior is calibrated with a label value of "0" or "1", the LSTM network input is the standardized periodic feature component, and the output is the maneuvering state label value.

9. The method according to claim 8, characterized in that The cross entropy loss function is defined as: , where For the The true label of the training sample corresponds to 0 or 1. For the model The probability of a sample being predicted as the corresponding maneuver label.

10. The method according to claim 8, characterized in that In step 3, Normalized value of the feature sequence at the moment As the input of the long short-term memory neural network, after the time series processing of the LSTM layer, the fully connected layer obtains the corresponding processing results, and then uses the softmax layer for classification. Finally, the classification layer outputs the result label of the maneuver detection, "0" represents that the target has not performed orbital maneuvers, or "1" represents that the target has performed orbital maneuvers.

Images