Abnormal multi-dimensional feature extraction and intelligent identification method for low-orbit satellite orbit determination link

Abstract

This application relates to a method for multi-dimensional feature extraction and intelligent identification of anomalies in low-Earth orbit (LEO) satellite orbit determination links. The method includes: extracting features from preprocessed signals and observation data from three dimensions—signal layer, data layer, and geometric layer—using a multi-dimensional feature extraction module to construct a seven-dimensional feature vector; building a differentiated fingerprint database representing different anomaly mechanisms based on the seven-dimensional feature vector; classifying and identifying anomaly types using a pre-trained lightweight hybrid intelligent classification model, and outputting the identified anomaly type and confidence level; and outputting the anomaly type and confidence level to the LEO satellite orbit determination system for adaptive fault-tolerant decision-making. This method overcomes the fundamental dependence of traditional precise orbit determination methods on a friendly, stable, and continuous observation environment, as well as the inherent defects of traditional robust filtering techniques, such as detection lag, unclear diagnosis, and limited strategy.

Description

Technical Field

[0001] This application relates to the field of satellite navigation signal processing and fault diagnosis technology, and in particular to a method for extracting and intelligently identifying anomalies in low-Earth orbit satellite orbit determination links. Background Technology

[0002] Precise orbit determination of low-Earth orbit satellites relies heavily on external observation data, such as that from Global Navigation Satellite Systems (GNSS). In the context of navigation countermeasures, GNSS signals are vulnerable to attacks such as intelligent deception and suppression interference, leading to systematic and covert anomalies in the observation data.

[0003] Traditional methods primarily rely on residuals after orbit determination for anomaly detection, which suffers from significant latency and cannot accurately distinguish anomaly types. Existing technologies such as RAIM (Receiver Autonomous Integrity Monitoring) mainly address geometric configuration issues and struggle to detect cooperative deception that does not disrupt geometric relationships; while robust filtering alone can only handle gross errors and is insensitive to low-amplitude, slowly varying deception. Therefore, there is an urgent need for a technology capable of early, accurate, and multi-dimensional anomaly detection and intelligent identification in the observation data link before orbit determination calculations, providing a prerequisite for subsequent fault-tolerant processing. Summary of the Invention

[0004] Therefore, it is necessary to provide a method for extracting and intelligently identifying anomalies in low-Earth orbit satellite orbit determination links that can overcome the fundamental dependence of traditional precision orbit determination methods on a friendly, stable, and continuous observation environment, as well as the inherent defects of traditional robust filtering techniques such as detection lag, unclear diagnosis, and single strategy.

[0005] A method for extracting and intelligently identifying anomalies in low-Earth orbit satellite orbit determination links, the method comprising:

[0006] The onboard processing unit receives GNSS observation signals from the low-Earth orbit satellite orbit determination link, preprocesses the GNSS observation signals, and obtains preprocessed signals and observation data. Based on the multi-dimensional feature extraction module, features are extracted from the preprocessed signal and observation data from three dimensions: signal layer, data layer, and geometric layer. A seven-dimensional feature vector is constructed, and a differentiated fingerprint database representing different anomaly mechanisms is built based on the seven-dimensional feature vector. A pre-trained lightweight hybrid intelligent classification model is used to classify and identify anomaly types in a seven-dimensional feature vector, and the identified anomaly type and confidence level are output. The anomaly type and confidence level are output to the low-Earth orbit satellite orbit determination system for adaptive fault-tolerant decision-making.

[0007] The aforementioned method for multi-dimensional feature extraction and intelligent identification of anomalies in low-Earth orbit satellite orbit determination links first collects and preprocesses GNSS observation signals, enabling anomaly detection without waiting for orbit determination calculations. This completely changes the lag problem of traditional methods relying on residual detection after orbit determination, achieving early anomaly perception and no longer being limited to friendly and stable observation environments. Secondly, the multi-dimensional feature extraction module jointly constructs a seven-dimensional feature vector from the signal layer, data layer, and geometric layer, forming a differentiated fingerprint database. The signal layer captures physical distortions such as carrier-to-noise ratio and correlation peaks; the data layer reflects statistical deviations such as pseudorange and Doppler; and the geometric layer relies on satellite observation geometric configuration verification. These three types of features comprehensively characterize different anomaly mechanisms, solving the diagnostic pain point of traditional single-feature or robust filtering methods being unable to distinguish anomaly types. Then, a lightweight hybrid intelligent classification model is set up, using a weighted fusion of LightGBM and 1D-CNN to capture both global statistical correlations of features and uncover local temporal patterns. It also supports online learning based on elastic weight consolidation to adapt to unknown anomalies. Compared to traditional single robust algorithms or models, it can flexibly cope with complex interference and novel anomalies, breaking the limitations of single strategies. Finally, the anomaly type and confidence level are output to the orbit determination system, providing a clear basis for the adaptive adjustment of the filtering strategy. This enables the orbit determination system to actively tolerate faults rather than passively respond, and it no longer relies on a continuous and stable observation environment. This significantly improves the autonomous survivability and orbit determination accuracy of the low-Earth orbit constellation in complex electromagnetic environments. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a method for extracting and intelligently identifying anomalies in a low-Earth orbit satellite orbit determination link, as described in one embodiment. Figure 2 This is a schematic diagram of a multi-dimensional feature extraction module in one embodiment; Figure 3 This is a schematic diagram of the intelligent anomaly recognition process in one embodiment; Figure 4 This is a schematic diagram of the on-orbit operation process and timing control in another embodiment. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0010] In one embodiment, such as Figure 1 As shown, a method for extracting and intelligently identifying anomalies in low-Earth orbit satellite orbit determination links is provided, including the following steps: Step 102: Receive GNSS observation signals from the low-Earth orbit satellite orbit determination link using the onboard processing unit, preprocess the GNSS observation signals, and obtain preprocessed signals and observation data.

[0011] The system receives raw intermediate frequency (IF) signals or baseband sampled signals from a GNSS receiver, along with demodulated observation data, including pseudorange, carrier phase, Doppler shift, and carrier-to-noise ratio. Simultaneously, it receives bidirectional ranging values ​​and time synchronization information from the inter-satellite link terminal via a bus. The preprocessing mainly involves filtering, amplifying, and performing analog-to-digital conversion on the received raw signals to remove some noise. The observation data undergoes format conversion and synchronization processing to prepare for subsequent feature extraction.

[0012] Step 104: Based on the multi-dimensional feature extraction module, feature extraction is performed on the preprocessed signal and observation data from three dimensions: signal layer, data layer, and geometric layer. A seven-dimensional feature vector is constructed, and a differentiated fingerprint database representing different anomaly mechanisms is built based on the seven-dimensional feature vector.

[0013] The multi-dimensional feature extraction module is implemented by a high-performance digital signal processor (DSP), which completes multi-dimensional feature extraction through parallel computing to ensure real-time performance. The signal layer features focus on the physical properties of the signal itself, the data layer features focus on the statistical regularity of the observed data, and the geometric layer features rely on the geometric configuration of satellite observations. The three-dimensional features work together to construct an anomaly fingerprint database, which can accurately characterize the unique mechanisms of different anomalies. The seven-dimensional feature vector is a high-discrimination input that integrates the three types of features, providing a comprehensive basis for intelligent recognition and solving the problem that traditional single features are insufficient for representing complex anomalies.

[0014] Step 106: Use a pre-trained lightweight hybrid intelligent classification model to classify and identify the anomaly types of the seven-dimensional feature vector, and output the identified anomaly types and confidence levels.

[0015] The lightweight hybrid intelligent classification model is a weighted fusion model of the LightGBM main classifier and the 1D-CNN auxiliary classifier. After INT8 quantization, it is adapted to the on-board embedded AI processor, balancing recognition accuracy and computational efficiency. It is pre-trained based on training data generated by the high-fidelity simulation software GSTAR, and injects anomalies such as spoofing interference, suppression interference, ionospheric scintillation, and receiver failure to generate tens of millions of labeled seven-dimensional feature samples. After model training, it is solidified into a binary engine file and burned to on-board storage. The anomaly types include five categories: normal, suppression interference, generative spoofing, ionospheric scintillation, and receiver failure. The confidence level reflects the reliability of the recognition results and provides a reference for subsequent fault-tolerant decisions.

[0016] Step 108: Output the anomaly type and confidence level to the low-Earth orbit satellite orbit determination system for adaptive fault-tolerant decision-making.

[0017] The adaptive fault-tolerant decision-making orbit determination system adjusts the filtering strategy and parameters based on the identification results. The orbit determination system adaptively adjusts the orbit determination filtering strategy and parameters based on the received anomaly type and confidence level, such as removing abnormal observation data and selecting more robust filtering algorithms. This solves the problem of inappropriate strategies for dealing with complex interference in traditional robust algorithms and improves the stability and accuracy of low-Earth orbit satellite orbit determination.

[0018] The aforementioned method for multi-dimensional feature extraction and intelligent identification of anomalies in low-Earth orbit satellite orbit determination links first collects and preprocesses GNSS observation signals, enabling anomaly detection without waiting for orbit determination calculations. This completely changes the lag problem of traditional methods relying on residual detection after orbit determination, achieving early anomaly perception and no longer being limited to friendly and stable observation environments. Secondly, the multi-dimensional feature extraction module jointly constructs a seven-dimensional feature vector from the signal layer, data layer, and geometric layer, forming a differentiated fingerprint database. The signal layer captures physical distortions such as carrier-to-noise ratio and correlation peaks; the data layer reflects statistical deviations such as pseudorange and Doppler; and the geometric layer relies on satellite observation geometric configuration verification. These three types of features comprehensively characterize different anomaly mechanisms, solving the diagnostic pain point of traditional single-feature or robust filtering methods being unable to distinguish anomaly types. Then, a lightweight hybrid intelligent classification model is set up, using a weighted fusion of LightGBM and 1D-CNN to capture both global statistical correlations of features and uncover local temporal patterns. It also supports online learning based on elastic weight consolidation to adapt to unknown anomalies. Compared to traditional single robust algorithms or models, it can flexibly cope with complex interference and novel anomalies, breaking the limitations of single strategies. Finally, the anomaly type and confidence level are output to the orbit determination system, providing a clear basis for the adaptive adjustment of the filtering strategy. This enables the orbit determination system to actively tolerate faults rather than passively respond, and it no longer relies on a continuous and stable observation environment. This significantly improves the autonomous survivability and orbit determination accuracy of the low-Earth orbit constellation in complex electromagnetic environments.

[0019] In one embodiment, a multi-dimensional feature extraction module extracts features from the preprocessed signal and observation data from three dimensions: signal layer, data layer, and geometric layer, constructing a seven-dimensional feature vector, including: The carrier-to-noise ratio anomaly characteristics, correlation peak distortion characteristics, and phase continuity characteristics are calculated at the signal layer. The pseudorange jump detection statistics, Doppler consistency verification error, and comprehensive signal quality index are calculated at the data layer. Calculate the geometric consistency check quantity at the geometric layer; A seven-dimensional feature vector is constructed based on carrier-to-noise ratio anomaly characteristics, correlation peak distortion characteristics, phase continuity characteristics, pseudorange jump detection statistics, Doppler consistency verification error, signal quality comprehensive index, and geometric consistency test quantity.

[0020] Specifically, the multi-dimensional feature extraction module employs parallel processing during feature extraction, simultaneously calculating features at the signal, data, and geometric layers. This significantly improves feature extraction efficiency and meets the requirements of real-time onboard processing. Signal layer features directly reflect the physical characteristics of GNSS signals. Carrier-to-noise ratio anomaly features capture abnormal fluctuations in signal energy, correlation peak distortion features characterize signal waveform distortion caused by deception interference, and phase continuity features detect carrier phase jumps. Data layer features are based on the statistical regularities of observed data. Pseudorange jump detection statistics identify abrupt changes in pseudorange data, Doppler consistency verification errors compare the differences between observed and theoretical Doppler, and comprehensive signal quality indicators integrate multi-dimensional low-level features. Geometric layer features rely on satellite observation geometry, with geometric consistency verification reflecting the residual distribution of multi-satellite observations. The seven-dimensional feature vector comprehensively covers the representation of anomalies from different dimensions. The constructed differentiated fingerprint database can accurately distinguish different anomaly mechanisms, solving the problem of insufficient ability of single features to identify complex anomalies and providing high-discrimination input for subsequent intelligent classification.

[0021] By extracting features from three different dimensions, anomalies in the orbit determination link can be comprehensively captured, avoiding the limitations of single-dimensional feature extraction. This results in a seven-dimensional feature vector with higher discriminative power, providing a more accurate basis for subsequent anomaly identification. For example, signal layer features mainly reflect anomalies in the quality and morphology of the signal itself, data layer features mainly reflect anomalies in the consistency and stability of the observed data, and geometric layer features focus on geometric configuration anomalies in satellite observations. The combination of these three features can cover a variety of anomalies that may occur in the low-Earth orbit satellite orbit determination link.

[0022] In one embodiment, the carrier-to-noise ratio anomaly characteristics, correlation peak distortion characteristics, and phase continuity characteristics are calculated at the signal layer, including: The carrier-to-noise ratio anomaly characteristics, correlation peak distortion characteristics, and phase continuity characteristics are calculated at the signal layer as follows:

[0023]

[0024]

[0025] in, This is a characteristic of abnormal carrier-to-noise ratio. and The mean and standard deviation within the sliding window. Carrier-to-noise ratio, It is a sliding window. This refers to the distortion characteristics of the relevant peaks. For the actual relevant peak shape, For the ideal correlation peak shape, The number of local maxima. For the actual relevant peak shape, For the ideal correlation peak shape, This represents the actual correlation peak skewness. For phase observations, The standard deviation of the second-order difference, i.e., the phase continuity characteristic. The length of the carrier phase observation sequence.

[0026] In one embodiment, the pseudorange jump detection statistic, Doppler consistency verification error, and comprehensive signal quality index are calculated at the data layer, including: The pseudorange jump detection statistic, Doppler consistency verification error, and comprehensive signal quality index are calculated at the data layer.

[0027]

[0028]

[0029] in, The actual observed pseudorange at the current epoch. Predict the pseudorange for the current epoch. For the prior standard deviation of pseudorange measurement, The pseudorange was actually observed in the previous epoch. The rate of change of pseudorange For epoch time intervals, For the epochal period, To observe the Doppler frequency shift, The calculated theoretical value, The relative velocity between satellites This is the unit vector between the satellite and the user. The standard deviation of Doppler observation noise. At the speed of light, The center frequency of the GNSS signal. This is the statistic for pseudorange jump detection. For Doppler consistency verification error, As a comprehensive indicator of signal quality, This represents the carrier-to-noise ratio.

[0030] In one embodiment, calculating the geometric consistency check metric at the geometry layer includes: The geometric consistency check metric is calculated at the geometric layer.

[0031] in, To observe the residual vector, For the observation matrix, To observe the noise covariance matrix, Number of visible satellites It follows an F-distribution, i.e., the degree of freedom parameter.

[0032] In one embodiment, the pre-trained lightweight hybrid intelligent classification model is a weighted fusion model of a LightGBM main classifier and a 1D-CNN auxiliary classifier. This pre-trained lightweight hybrid intelligent classification model is used to classify and identify anomaly types from the seven-dimensional feature vector, outputting the identified anomaly type and its confidence level, including: The seven-dimensional feature vector is input into the LightGBM main classifier, and the output is the first probability distribution of the sample belonging to each anomaly type. The seven-dimensional feature vector is input into the 1D-CNN auxiliary classifier, and the output is the second probability distribution of the sample belonging to each anomaly type. The first probability distribution and the second probability distribution are weighted and fused to obtain the final anomaly type probability distribution; The anomaly type and its corresponding confidence level are determined based on the final anomaly type probability distribution.

[0033] Specifically, the LightGBM main classifier excels at capturing global statistical correlations between features, adapting to the overall pattern mining of seven-dimensional features; the 1D-CNN auxiliary classifier can capture local patterns and temporal correlations in feature vectors, compensating for the main classifier's insufficient perception of fine-grained features. The weighted fusion of these two classes achieves complementary advantages, solving the problem of insufficient accuracy of a single model in identifying complex anomalies. Model training data is generated through GSTAR simulation, setting the orbital altitude, injecting various anomalies conforming to physical laws, generating labeled samples after feature extraction, and quantizing them into INT8 format after training to meet the low-power inference requirements of the onboard NPU, ensuring real-time performance and reliability during on-orbit operation.

[0034] In one embodiment, the loss function of the LightGBM main classifier is:

[0035] in, This is the set of learnable parameters for the LightGBM model. The total number of training samples, C This represents the total number of exception types. For the first i The sample belongs to the first c The true label of the class, For the first i The predicted probability that a sample belongs to class c. This is a complexity control parameter for leaf node splitting. T This represents the total number of leaf nodes in the current decision tree. The penalty coefficient is...w The output weight vector for the decision leaf nodes; The 1D-CNN auxiliary classifier is trained using a multi-class cross-entropy loss function, the loss function being...

[0036] in, The total number of training samples, For the first i The sample belongs to the first c The true label of the class, The first prediction of the 1D-CNN model i The sample belongs to the first c The probability of a class.

[0037] The 1D-CNN auxiliary classifier captures deeper local patterns and temporal correlations in the feature vectors.

[0038] In one embodiment, the LightGBM master classifier selects the optimal splitting feature using a feature splitting gain formula; the feature splitting gain formula is as follows:

[0039] in, For the loss function on the th i The first gradient of each sample For the loss function on the th i The second gradient of each sample. , These are the sample sets of the left and right child nodes after the split, respectively. The penalty coefficient is... This is a complexity control parameter for leaf node splitting.

[0040] Specifically, the gain is determined by calculating the change in loss before and after splitting, and the feature with the largest gain is prioritized for splitting. This allows for the accurate discovery of the feature dimension that contributes most to anomaly detection, improving the model's recognition accuracy. The first and second gradients are obtained by differentiating the loss function, and a regularization term is used to balance the splitting gain with model complexity, avoiding overfitting caused by excessive splitting. This splitting strategy enables LightGBM to efficiently utilize the discriminative information of seven-dimensional features, quickly locate key feature dimensions, and adapt to the needs of lightweight inference on satellites.

[0041] In one embodiment, the weighted fusion method is as follows:

[0042] in, For fused samples x Belongs to the c The probability of a class The probability output by the LightGBM main classifier. The probability output by the 1D-CNN auxiliary classifier. For fusion weights and 0≤ ≤1.

[0043] Specifically, this formula captures the probability of fault categories in the observed data. The fused probability distribution combines the identification results of the two models, reducing the risk of misjudgment by a single model and improving the accuracy and robustness of anomaly identification, especially for complex anomalies with blurred boundaries, where it has a better classification effect.

[0044] In one embodiment, the lightweight hybrid intelligent classification model supports online learning based on elastic weight reinforcement to adapt to unknown anomaly types. The optimization objective function for online learning is:

[0045] in, For the loss function on the new task, The penalty coefficient is... for Fisher The diagonal elements of the information matrix are used to constrain important parameters. Changes, For the current parameters of the model, These are the optimal parameters for the model on the old task.

[0046] Specifically, this online learning mechanism is achieved through... Fisher Information Matrix Mark key parameters and penalty coefficients in old tasks. By controlling the magnitude of parameter updates and constraining changes in key parameters when learning unknown anomaly patterns, the model avoids forgetting previously learned knowledge. When a high-confidence unknown anomaly type is detected during on-orbit operation, the model can incrementally update based on newly acquired labeled samples without retraining the entire model, thus adapting to the emergence of new anomalies in navigation adversarial scenarios. This mechanism enhances the model's adaptability, solves the problem of traditional models struggling to cope with unknown anomalies, and improves the long-term robustness of the system in complex electromagnetic environments by incrementally updating model parameters while avoiding the forgetting of previously learned knowledge.

[0047] In a specific embodiment, the DSP obtains the current epoch. t baseband signal and Observations. By maintaining a length of L=100 A circular buffer for historical sequences is used to calculate the sample mean of the buffer. and sample standard deviation : (1) (2) Calculate the normalized bias for the current epoch: (3) To avoid redundant calculations, the DSP uses a recursive algorithm for updating. and When new data The arrival of old data When removing: (4) (5) in, If the sum of squares is , then .

[0048] Simultaneously, by extracting correlation peak distortion features to quantify the difference between the actual correlation peak and the ideal peak, the output of the internal correlator of the GNSS receiver is... ,in This is a local pseudo-random code copy. The DSP reads it directly from the receiver. At several chip offsets The value at that location. Store an ideal correlation peak. , Let be the chip width. Then, calculate the correlation peak distortion characteristics: (6) In the formula, —Euclidean distance; — represents the skewness term. for The number of sampling points, and Its mean and standard deviation; —Multi-peak penalty item, This represents the number of peaks obtained through the local maximum search algorithm; Transitions are detected using higher-order differences in carrier phase observations. A carrier phase is maintained. A short sequence of length . Calculate the second-order difference sequence Calculate the standard deviation of this second-order difference sequence: (7) In the formula, —The mean of a second-order difference sequence; A simple uniform velocity model is used for prediction. Let the state of the previous epoch be the position. and speed This can be provided by the GSTAR orbit determination filter. Therefore, the predicted current epoch position is: For the first GNSS satellites, predicted pseudorange: ,in The satellite's position is calculated from the ephemeris. This refers to the receiver clock bias.

[0049] Further, calculate the test statistic: (8) In the formula, —The prior standard deviation of the distance measurement is generally set to 0.3m; Furthermore, the Doppler consistency verification error is extracted, and the geometrically calculated Doppler is compared with the observed Doppler. The theoretical Doppler calculation includes: (9) In the formula, —GNSS signal center frequency, GPS L1 is 1575.42J=MHz; —speed of light ; —The velocity vector of the i-th GNSS satellite; —The velocity vector of the low-Earth orbit satellite; —The unit line-of-sight vector from the low-Earth orbit satellite to the i-th GNSS satellite; Further, extract features: (10) In the formula, —The standard deviation of Doppler observation noise can be set to 0.5Hz; Furthermore, extract comprehensive signal quality indicators: (11) Receiver Autonomous Integrity Monitoring (RAIM) based on least-squares residuals. If there is currently a... m Visible by several satellites. Equation construction: ,in This is the pseudorange observation vector. This is a geometric matrix; the first three columns are the line-of-sight vectors, and the fourth column is... , The status is pending evaluation. It is noise.

[0050] Furthermore, least squares estimation is performed using all m satellites: Calculate the residual vector Calculate the unit weighted variance: For the i-th satellite, its observations are discarded, and the least squares solution is performed again using the remaining m-1 satellites to obtain the sum of squared residuals. .

[0051] Furthermore, construct the test statistic: (12) This statistic reflects the contribution of the i-th satellite to the overall residual. The higher the value, the greater the likelihood of anomalies in the satellite observations. The final value is... This is a characteristic of that epoch.

[0052] Figure 3 The training and solidification of a lightweight hybrid intelligent model are described. The training data simulation uses the high-fidelity simulation software GSTAR to simulate GNSS observation data from low-Earth orbit satellites under different scenarios, with the orbital altitude set to 500-800 km. A key step is the injection of physically consistent anomalies, categorized as deception interference, suppression interference, ionospheric scintillation, and faults. Furthermore, the simulated injection of deception interference is not simply adding a bias, but rather using a simulated generative deception device, the signal of which is: (13) In the formula, —Delay deception quantity; —Phase deception quantity; Both of these parameters change slowly to simulate a real deception process.

[0053] Furthermore, simulations are used to suppress interference by adding Gaussian white noise or narrowband interference signals in the time or frequency domains. , making And control the interference ratio.

[0054] Furthermore, it injects flickering effects and fault interruptions.

[0055] Furthermore, features are extracted and labels are set. The S100 step of this application is run on the massive amount of data generated by the simulation to generate tens of millions of seven-dimensional feature vector samples, each with a precise category label—normal, deception, suppression, flickering, and fault.

[0056] Next, model training is performed. For LightBGM, the lightgbm library in Python is used. Key parameter settings are shown in Table 1: Table 1

[0057] For 1D-CNN, PyTorch is used with Adam as the optimizer and a learning rate of 1e-4.

[0058] Validate the performance of the two models on independent test sets. Determine the optimal fusion weights. The fused model is treated as a whole and then quantized after training. The weights and activation values ​​are quantized from FP32 to INT8.

[0059] This application is based on pure simulation. If the requirements of actual measurement and computing power are met, the quantized model parameters and network structure definition files can be compiled into a binary engine file that can run efficiently on a specific NPU using the toolchain provided by the chip manufacturer. This engine file is then burned into the eMMC storage of the on-board processing unit as part of the firmware.

[0060] Figure 4 The on-orbit operation process and timing control of this invention are described. To ensure real-time performance, the system employs strict timing control. The timing within one processing cycle (e.g., 1 second) is shown in Table 2: Table 2

[0061] This application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it can implement the GNSS matched spectrum interference suppression method based on the polarity cut-off algorithm described in any of the above embodiments.

[0062] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0063] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0064] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for extracting and intelligently identifying anomalies in low-Earth orbit satellite orbit determination links, characterized in that, The method includes: The onboard processing unit receives GNSS observation signals from the low-Earth orbit satellite orbit determination link, and preprocesses the GNSS observation signals to obtain preprocessed signals and observation data. The multi-dimensional feature extraction module extracts features from the preprocessed signal and observation data from three dimensions: signal layer, data layer, and geometric layer, and constructs a seven-dimensional feature vector. Based on the seven-dimensional feature vector, a differentiated fingerprint database representing different anomaly mechanisms is constructed. The seven-dimensional feature vector is classified and identified using a pre-trained lightweight hybrid intelligent classification model, and the identified anomaly type and confidence level are output. The anomaly type and confidence level are output to the low-Earth orbit satellite orbit determination system for adaptive fault-tolerant decision-making.

2. The method according to claim 1, characterized in that, The preprocessed signal and observation data are subjected to feature extraction from three dimensions—signal layer, data layer, and geometric layer—using a multi-dimensional feature extraction module to construct a seven-dimensional feature vector, including: The carrier-to-noise ratio anomaly characteristics, correlation peak distortion characteristics, and phase continuity characteristics are calculated at the signal layer. The pseudorange jump detection statistics, Doppler consistency verification error, and comprehensive signal quality index are calculated at the data layer. Calculate the geometric consistency check quantity at the geometric layer; A seven-dimensional feature vector is constructed based on the carrier-to-noise ratio anomaly characteristics, correlation peak distortion characteristics, phase continuity characteristics, pseudorange jump detection statistics, Doppler consistency verification error, signal quality comprehensive index, and geometric consistency test quantity.

3. The method according to claim 2, characterized in that, The signal layer is used to calculate carrier-to-noise ratio anomaly characteristics, correlation peak distortion characteristics, and phase continuity characteristics, including: The carrier-to-noise ratio anomaly characteristics, correlation peak distortion characteristics, and phase continuity characteristics are calculated at the signal layer as follows: in, This is a characteristic of abnormal carrier-to-noise ratio. and The mean and standard deviation within the sliding window. Carrier-to-noise ratio, It is a sliding window. This refers to the distortion characteristics of the relevant peaks. For the actual relevant peak shape, For the ideal correlation peak shape, The number of local maxima. For the actual relevant peak shape, For the ideal correlation peak shape, This represents the actual correlation peak skewness. For phase observations, The standard deviation of the second-order difference, i.e., the phase continuity characteristic. The length of the carrier phase observation sequence.

4. The method according to claim 2, characterized in that, The pseudorange jump detection statistics, Doppler consistency verification error, and comprehensive signal quality indicators are calculated at the data layer, including: The pseudorange jump detection statistic, Doppler consistency verification error, and comprehensive signal quality index are calculated at the data layer. in, The actual observed pseudorange at the current epoch. Predict the pseudorange for the current epoch. For the prior standard deviation of pseudorange measurement, The pseudorange was actually observed in the previous epoch. The rate of change of pseudorange For epoch time intervals, For the epochal period, To observe the Doppler frequency shift, The calculated theoretical value, The relative velocity between satellites This is the unit vector between the satellite and the user. The standard deviation of Doppler observation noise. At the speed of light, The center frequency of the GNSS signal. This is the statistic for pseudorange jump detection. For Doppler consistency verification error, As a comprehensive indicator of signal quality, This represents the carrier-to-noise ratio.

5. The method according to claim 2, characterized in that, Calculate geometric consistency check quantities at the geometric layer, including: The geometric consistency check metric is calculated at the geometric layer. in, To observe the residual vector, For the observation matrix, To observe the noise covariance matrix, Number of visible satellites It follows an F-distribution, i.e., the degree of freedom parameter.

6. The method according to claim 1, characterized in that, The pre-trained lightweight hybrid intelligent classification model is a weighted fusion model of a LightGBM main classifier and a 1D-CNN auxiliary classifier. This model is used to classify and identify anomaly types in the seven-dimensional feature vector, outputting the identified anomaly types and their confidence levels, including: The seven-dimensional feature vector is input into the LightGBM main classifier, and the output is the first probability distribution of the sample belonging to each anomaly type. The seven-dimensional feature vector is input into a 1D-CNN auxiliary classifier, and the output is the second probability distribution of the sample belonging to each anomaly type. The first probability distribution and the second probability distribution are weighted and fused to obtain the final anomaly type probability distribution; The anomaly type and its corresponding confidence level are determined based on the final anomaly type probability distribution.

7. The method according to claim 6, characterized in that, The loss function of the LightGBM main classifier is: in, This is the set of learnable parameters for the LightGBM model. The total number of training samples, C This represents the total number of exception types. For the first i The sample belongs to the first c The true label of the class, For the first i The predicted probability that a sample belongs to class c. This is a complexity control parameter for leaf node splitting. T This represents the total number of leaf nodes in the current decision tree. The penalty coefficient is... w The output weight vector for the decision leaf nodes; The 1D-CNN auxiliary classifier is trained using a multi-class cross-entropy loss function, which is: in, The total number of training samples, For the first i The sample belongs to the first c The true label of the class, The first prediction of the 1D-CNN model i The sample belongs to the first c The probability of a class.

8. The method according to claim 6, characterized in that, The LightGBM master classifier selects the optimal splitting feature using a feature splitting gain formula; the feature splitting gain formula is: in, For the loss function on the th i The first gradient of each sample For the loss function on the th i The second gradient of each sample. , These are the sample sets of the left and right child nodes after the split, respectively. The penalty coefficient is... This is a complexity control parameter for leaf node splitting.

9. The method according to claim 6, characterized in that, The weighted fusion method is as follows: in, For fused samples x Belongs to the c The probability of a class The probability output by the LightGBM main classifier. The probability output by the 1D-CNN auxiliary classifier. For fusion weights and 0≤ ≤1.

10. The method according to claim 6, characterized in that, The lightweight hybrid intelligent classification model supports online learning based on elastic weight reinforcement to adapt to unknown anomaly types. The optimization objective function of the online learning is: in, For the loss function on the new task, The penalty coefficient is... for Fisher The diagonal elements of the information matrix, For the current parameters of the model, These are the optimal parameters for the model on the old task.

Images