Method and apparatus for extracting multi-layer physical features of gnss vector tracking receiver

Abstract

The application relates to a GNSS vector tracking receiver multi-layer physical feature extraction method and device. The method comprises the following steps: acquiring a received signal of a GNSS vector tracking receiver; performing baseband correlation processing on the received signal to obtain multi-channel correlator output data, and extracting baseband correlation layer features representing correlation peak shapes based on the multi-channel correlator output data; extracting navigation filtering layer features representing closed-loop running states of a vector tracking loop based on output data of an extended Kalman filtering module of the vector tracking loop; extracting common mode fingerprint layer features representing cross-channel linkage characteristics based on joint statistical results of all satellite channels; and performing statistical quantization and standardization processing on the baseband correlation layer features, the navigation filtering layer features and the common mode fingerprint layer features according to fixed-length non-overlapping windows to obtain multi-layer physical features corresponding to each epoch. By using the method, the feature distinguishability in a concealed smart scene can be improved without increasing the hardware cost of the receiver.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and apparatus for extracting multi-layer physical features of a GNSS vector tracking receiver. Background Technology

[0002] Global Navigation Satellite System (GNSS) provides continuous and unified positioning, navigation, and timing (PNT) services to various users, and is widely used in critical fields such as transportation, power communications, unmanned systems, financial timing, and national defense. However, civilian GNSS signals reach the ground with extremely low power, and their signal structure is publicly known, making it difficult for receivers to directly authenticate genuine satellite signals from spoofed signals based solely on the signal architecture. Therefore, GNSS receivers are vulnerable to smart attacks.

[0003] Unlike suppression jamming, smart attacks typically do not manifest primarily as significant power anomalies or signal loss. Attackers can transmit spoofed signals highly similar to real satellite signals and, through power control, code phase alignment, carrier dragging, and navigation message synchronization, cause the receiver to output erroneous PVT calculation results while maintaining apparent normal tracking. With the development of software-defined radio platforms, programmable signal simulators, and portable smart sources, the threshold for implementing smart attacks has been continuously lowered, and attack forms have evolved from high-power, abrupt smart attacks to more covert forms such as power matching, slow dragging, and multi-channel coordinated takeover. Therefore, achieving high detection probability, low false alarm rate, and interpretable GNSS smart detection without significantly increasing receiver hardware costs has become a key issue in receiver security enhancement.

[0004] Traditional GNSS receivers often employ a Scalar Tracking Loop (STL), where each satellite channel independently acquires and tracks the satellite, lacking information sharing between channels. In contrast, a Vector Tracking Loop (VTL) deeply couples signal tracking with navigation calculations through an extended Kalman filter, mapping multi-satellite observations to a unified receiver state estimate and further feeding them back to the local reference signal generation process for each channel. This architecture not only improves tracking continuity in weak signal, obstructed, and dynamic scenarios but also provides a system-level observation window within the receiver for agile detection. Once agile biases enter the measurement domain, they can propagate to the global state through filter gain, leaving observable traces in correlation peak shapes, innovation statistics, feedback control variables, and cross-channel coupling relationships. Therefore, fully utilizing the internal physical constraints and multi-channel coupling information of the VTL is an important direction for constructing low-cost, deployable GNSS agile detection methods.

[0005] In recent years, machine learning and deep learning methods have been applied to GNSS smart detection. Methods such as LSTM, convolutional neural networks, Transformers, and generative models can learn nonlinear discriminative relationships from multidimensional receiver observations, achieving performance superior to traditional threshold detection on some public datasets and experimental scenarios. However, most existing data-driven methods still use STL scenarios, flattened statistical vectors, or ordinary time series as inputs, and model design mainly focuses on improving classification accuracy. This design pattern cannot adapt to the closed-loop working mechanism of vector tracking receivers, making it difficult to uncover the internal signal variation patterns and inter-channel linkage features. In concealed smart scenarios such as power matching and slow dragging, not only does the detection accuracy decrease, but the output results also lack corresponding physical basis. In addition, detection of vector tracking receivers faces a series of inherent technical obstacles: the internal dynamic computational parameters generated during equipment operation cannot be directly obtained through conventional acquisition methods; the correlation between different satellite channels with changes in satellite geometry and receiver operating conditions is complex and difficult to quantify; and the multi-layered physical constraints formed by signal transmission and loop control, coupled with interference from smart signals, further weaken the discriminative power of effective features. In summary, how to overcome the technical bottlenecks in feature extraction, construct a feature extraction scheme adapted to vector tracking loops, and thereby improve the receiver's comprehensive detection capability for concealed and agile signals has become a core problem that urgently needs to be solved in this field. Summary of the Invention

[0006] Therefore, it is necessary to provide a method and apparatus for extracting multi-layer physical features of a GNSS vector tracking receiver to address the aforementioned technical problems.

[0007] A method for extracting multi-layer physical features from a GNSS vector tracking receiver, the method comprising: Acquire the received signal from the GNSS vector tracking receiver; Baseband correlation processing is performed on the received signal to obtain multi-channel correlator output data, and baseband correlation layer features characterizing the correlation peak morphology are extracted based on the multi-channel correlator output data. The output data of the extended Kalman filter module based on the vector tracking loop is used to extract navigation filter layer features that characterize the closed-loop operation state of the loop; the output data includes navigation state estimation, innovation residuals and feedback control data; Common-mode fingerprint layer features characterizing cross-channel linkage are extracted based on joint statistical results from all satellite channels; The features of the baseband correlation layer, navigation filter layer, and common mode fingerprint layer are statistically quantized and standardized using a fixed-length non-overlapping window to obtain the multi-layer physical features corresponding to each epoch.

[0008] A multi-layer physical feature extraction device for a GNSS vector tracking receiver, the device comprising: The signal acquisition module is used to acquire the received signal from the GNSS vector tracking receiver; The baseband correlation layer feature extraction module is used to perform baseband correlation processing on the received signal to obtain multi-channel correlator output data, and extract baseband correlation layer features characterizing the correlation peak shape based on the multi-channel correlator output data; The navigation filter layer feature extraction module is used to extract navigation filter layer features that characterize the closed-loop operation state of the extended Kalman filter module based on the output data of the vector tracking loop; the output data includes navigation state estimation, innovation residual and feedback control quantity data; The common-mode fingerprint layer feature extraction module is used to extract common-mode fingerprint layer features that characterize cross-channel linkage based on the joint statistical results of all satellite channels; The multi-layer physical feature output module is used to perform statistical quantization and standardization on the baseband correlation layer features, navigation filter layer features, and common mode fingerprint layer features using a fixed-length non-overlapping window to obtain the multi-layer physical features corresponding to each epoch.

[0009] The aforementioned GNSS vector tracking receiver multi-layer physical feature extraction method and apparatus, by constructing a hierarchical physical feature extraction architecture consisting of a baseband correlation layer, a navigation filter layer, and a common-mode fingerprint layer, can adapt to the closed-loop working mechanism of the vector tracking receiver. It covers the complete propagation link of interference from the baseband inlet to the filter closed loop and then to cross-channel diffusion. Specifically, by calling the internal computational data of the extended Kalman filter module to extract navigation filter layer features, it can uncover loop dynamic parameters that are unavailable through conventional acquisition methods. It utilizes the multi-channel coupling characteristics to amplify system-level anomalies caused by covert interference. Through joint statistical extraction of common-mode fingerprint layer features across all satellite channels, it can quantify and characterize the dynamic linkage relationship between channels, capture synchronization disturbances caused by single-source interference, and supplement the detection blind spots of single-channel features. Through statistical quantization and standardization processing using fixed-length non-overlapping windows, it can suppress instantaneous noise fluctuations, unify the scale of multi-dimensional feature dimensions, and ensure the stability and comparability of features. This invention can improve feature discrimination in covert and agile scenarios without increasing receiver hardware costs, providing reliable feature support for GNSS agile detection with high detection probability, low false alarm rate, and physical interpretability. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a multi-layer physical feature extraction method for a GNSS vector tracking receiver in one embodiment; Figure 2 This is a schematic diagram of the timing response of 15 features in one embodiment. Detailed Implementation

[0011] 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.

[0012] In one embodiment, such as Figure 1 As shown, a method for extracting multi-layer physical features of a GNSS vector tracking receiver is provided, including the following steps: Step 102: Obtain the received signal from the GNSS vector tracking receiver; Vector tracking receivers are receiver types that employ a vector tracking loop architecture. They deeply couple multi-satellite signal tracking with navigation state calculation through an extended Kalman filter. Each satellite channel shares the global navigation state and provides unified feedback control of the local reference signal, which distinguishes them from scalar tracking receivers where each channel operates independently.

[0013] Step 104: Perform baseband correlation processing on the received signal to obtain multi-channel correlator output data, and extract baseband correlation layer features that characterize the correlation peak morphology based on the multi-channel correlator output data.

[0014] Baseband correlation processing is the core processing step in a GNSS receiver's baseband, involving correlation operations between the received intermediate frequency (IF) signal and a locally generated pseudo-random code to achieve pseudo-code synchronization and signal energy accumulation. A multi-channel correlator is the correlation operation unit in the receiver corresponding to each visible satellite, typically configured with early, accurate, and late symmetrical correlation taps. Baseband correlation layer features are a set of features extracted from the baseband correlation output that reflect changes in the shape of the signal correlation peaks.

[0015] It is understandable that this step extracts relevant peak shape features from the very beginning of signal access, which can capture the baseband signal distortion caused in the early stage of smart signal injection. This is beneficial for identifying abnormal signals in the early stages of smart interference access, and at the same time provides basic data support for the baseband dimension of the subsequent multi-level feature system.

[0016] Step 106: Extract navigation filter layer features characterizing the closed-loop operating state of the vector tracking loop from the output data of the extended Kalman filter module. The output data includes navigation state estimation, innovation residuals, and feedback control quantity data.

[0017] The extended Kalman filter module is the core computational unit of the vector tracking loop, used to fuse multi-satellite observation data to achieve optimal estimation of the receiver's position, velocity, clock bias, and other navigation states. Navigation state estimation refers to the estimated results of core navigation parameters such as receiver position, velocity, clock bias, and clock drift output by the filter. Innovation residual refers to the difference between the actual measured value and the filter's predicted value, reflecting the degree of deviation between the measured and predicted results. Feedback control signals are control signals obtained by mapping the navigation state estimation and used to control the local code and carrier oscillator of each channel. Navigation filter layer features are a set of features extracted from the internal computational data of the filter, reflecting anomalies in the closed-loop operation state.

[0018] It is understandable that this step delves into the core filtering unit of the vector tracking loop to extract features, which can uncover dynamic parameters inside the loop that cannot be directly obtained by conventional acquisition methods. It fully utilizes the closed-loop coupling characteristics of vector tracking to amplify system-level anomalies caused by agile interference, which is beneficial to improving the distinguishability of features in concealed and agile scenarios.

[0019] Step 108: Extract common mode fingerprint layer features that characterize cross-channel linkage characteristics based on the joint statistical results of all satellite channels.

[0020] Joint statistics across all satellite channels refer to the approach of treating all visible satellite channel observation data as a whole for statistical analysis, rather than analyzing each channel independently, to uncover correlations between channels. Cross-channel linkage characteristics refer to the synchronous changes exhibited by each satellite channel in vector tracking scenarios due to sharing the same navigation state and receiver clock; single-source agility further amplifies this synchronous disturbance. Common-mode fingerprint layer features are a set of features extracted from multi-channel synchronous changes that reflect global common-mode disturbances.

[0021] It is understandable that this step extracts features from the multi-channel global correlation dimension, which can quantitatively characterize the linkage relationship between satellite channels as the geometric distribution and operating conditions change dynamically, capture cross-channel synchronization disturbances caused by single-source ingenuity, and help supplement the global common-mode anomaly dimension not covered by the baseband layer and filter layer, and improve the detection coverage dimension of the feature system.

[0022] Step 110: Perform statistical quantization and standardization on the baseband correlation layer features, navigation filter layer features, and common mode fingerprint layer features using a fixed-length non-overlapping window to obtain the multi-layer physical features corresponding to each epoch.

[0023] Fixed-length non-overlapping windows refer to dividing time-series data into non-overlapping segments with a fixed number of sampling points, with each segment undergoing independent statistical calculations. Statistical quantization refers to performing aggregated statistical calculations on the time-series feature data within the window to obtain stable feature values ​​at the window level. Standardization refers to scaling features based on the statistical characteristics of normal samples to eliminate the influence of differences in the dimensions of different features. Multi-layer physical features refer to a structured feature set covering three levels—baseband, filtering, and common mode—after unified processing.

[0024] It is understandable that this step performs unified window statistics and standardization on the three-layer features, which can suppress the impact of instantaneous noise fluctuations on feature stability, eliminate the dimensional differences between different physical quantities, and obtain a structured feature set that is time-aligned and scale-uniform, which is beneficial for subsequent detection models to directly call and process.

[0025] The aforementioned multi-layer physical feature extraction method for GNSS vector tracking receivers constructs a hierarchical physical feature extraction architecture comprising a baseband correlation layer, a navigation filter layer, and a common-mode fingerprint layer. This architecture adapts to the closed-loop operating mechanism of vector tracking receivers, covering the complete propagation link of interference from the baseband inlet to the filter closed loop and then to cross-channel diffusion. Specifically, by extracting navigation filter layer features using internal computational data from the extended Kalman filter module, loop dynamic parameters that are unavailable through conventional acquisition methods can be uncovered. The multi-channel coupling characteristics amplify system-level anomalies caused by covert interference. Joint statistical extraction of common-mode fingerprint layer features across all satellite channels quantifies the dynamic linkage between channels, captures synchronization disturbances caused by single-source interference, and supplements the detection blind spots of single-channel features. Statistical quantization and standardization using fixed-length non-overlapping windows suppresses instantaneous noise fluctuations, unifies the scale of multi-dimensional feature dimensions, and ensures the stability and comparability of features. This invention improves feature discrimination in covert and agile scenarios without increasing receiver hardware costs, providing reliable feature support for high-probability, low-false-alarm-rate, and physically interpretable GNSS agile detection.

[0026] In this embodiment, it is worth noting that the correlator output is composed of the true signal component, the agile signal component, and the noise term. Therefore, from the perspective of signal quality monitoring, agile interference mainly changes the statistical characteristics of the receiver's correlator in three ways: first, it changes the power and fluctuation mode of the Prompt branch, thereby affecting the robust carrier-to-noise ratio estimation; second, it disrupts the mirror symmetry relationship between Early and Late, thereby inducing correlation peak distortion skewness; and third, as the code separation continues to increase, it causes the composite correlation envelope to gradually evolve from a single peak to a double-peak or shoulder-peak structure. The above-mentioned baseband-level anomalies will further propagate along the vector tracking loop to the filter closed loop. In the scalar DLL, the equivalent noise bandwidth is usually given by the fixed loop filter coefficients. However, in the VDLL, the control quantity of the code NCO is generated by the unified feedback of the navigation state output by the EKF, so its equivalent noise bandwidth should be described by the vector transfer function of the code frequency closed loop. The VDLL equivalent noise bandwidth is not a fixed loop parameter, but is determined by the geometric mapping matrix. With EKF position-clock difference gain submatrix The time-dependent variables are jointly determined. When agile interference enters the pseudorange measurement, the innovation statistics and measurement noise model will change simultaneously, and further alter the... The numerical distribution of the noise causes abnormal fluctuations in the equivalent noise bandwidth of the vector tracking loop. The innovation is the core bridge connecting the anomaly in the measurement domain and the internal state changes of the filter. The effect of cleverness on the innovation has two meanings: first, it changes the instantaneous amplitude distribution of the innovation, increasing the normalized innovation norm; second, it disrupts the temporal structure of the innovation, which approximates white noise, causing statistical correlation between adjacent epochs. Based on the differentiated propagation characteristics of clever interference at the three levels of baseband correlation, closed-loop gain, and filter statistics, and to balance physical interpretability, hierarchical complementarity, and consistency in engineering implementation, this invention divides the detection features into three levels, constructing a corresponding multi-layered physical feature system.

[0027] In one specific embodiment, the present invention divides 15 smart detection features into three levels, as shown in Table 2. First, the baseband correlation layer features mainly characterize the distortion of correlation peak morphology, and are more sensitive to local correlation peak anomalies in the early stage of smart drag, thus reflecting the baseband observation changes caused by smart signal injection earlier. Second, the navigation filter layer features are used to characterize VDLL closed-loop gain changes, innovation statistical anomalies, and cross-channel consistency disruptions among multi-satellite observations, revealing the systematic dynamic response induced by smart bias entering the EKF from the filter estimation level. Finally, the baseband common-mode fingerprint layer features are geared towards single-source smart scenarios such as digital intermediate frequency injection, supplementing the characterization of common-mode perturbation characteristics of smart signals in multiple channels.

[0028] The agile detection features defined in Table 1 are extracted using a GNSS software receiver. Taking the ds2 agile scene in the TEXBAT dataset as an example, the temporal response of the 15 features is as follows: Figure 2 As shown, different levels of features exhibit differentiated change patterns before and after smart access, indicating that the constructed feature system can complementaryly characterize smart interference from three levels: baseband correlation, navigation filtering, and common-mode fingerprinting. The relevant parameter settings in the software are as follows: the receiver's coherent integration time and filter update period are still set to... Early-Late nominal interval is taken The multiple correlator uses 21 symmetrical taps. Furthermore, all window statistics use a fixed-length 50-point non-overlapping window instead of a continuous sliding window. Then the actual duration corresponding to a statistical window is . No. The epoch index set of each window is defined as

[0029] If the epoch-making sequence of a certain window is denoted as Then its window mean can be expressed as .

[0030] Table 1. Hierarchical Design of Smart Detection Features

[0031] Baseband related layer feature design (1) Robust carrier-to-noise ratio (BB1): The first Each channel in the historical era The Prompt power is

[0032] For each window The samples within the range are used to construct the second and fourth moments, i.e.

[0033]

[0034] When the signal and noise components are approximately independent, the instantaneous signal-to-noise ratio given by the M2 / M4 estimator can be expressed as:

[0035] Therefore, the main feature of BB1 is defined as

[0036] This means that when the smart signal enters the correlator, the effective power of the Prompt branch is no longer determined solely by the real signal and thermal noise, but is simultaneously superimposed with the smart component and the crossover component between the real and smart signals, thus leading to... and Simultaneous offset, ultimately manifested as Abnormal rise or fall.

[0037] (2) Correlation peak skewness coefficient (BB2): BB2 is used to model the overall shape of the correlation peak using the output of a 21-tap multicorrelator. Let the first... One channel is tapping. , calendar The relevant amplitude is denoted as In each window The time averaging of the envelope of the multiple correlators is performed internally.

[0038] After normalization, we can obtain Based on this, the weighted centroid and weighted variance are defined as follows:

[0039]

[0040] Therefore, the correlation peak skewness coefficient is defined as...

[0041] It can be seen that this statistic is essentially the third central moment of the normalized correlation envelope. Without cleverness, the correlation peaks approximately satisfy mirror symmetry on both sides of the Prompt, therefore... Approaching zero; however, the symmetry will be disrupted by a slow, dexterous pull that causes a unilateral acromion, unilateral stretching, or bi-peak offset, resulting in a significant deviation of the third moment from zero.

[0042] (3) Related side peak ratio (BB3): BB3 and BB2 share the same normalized envelope. It detects structures where the main peak and secondary peak coexist. If a certain tap satisfies...

[0043] This tap is considered to correspond to a local peak. Let the amplitudes of the highest and second-highest peaks be respectively... , The ratio of related secondary peaks is defined as follows:

[0044] This indicates that if both the true peak and the clever peak exist simultaneously within the current window, then... It will increase significantly; if the relevant envelope still maintains a single-peak structure, then It degenerates to zero.

[0045] 3.3 Navigation Filter Layer Feature Design (1) Vector DLL equivalent noise bandwidth (NF1): In the vector tracking loop, the code loop equivalent noise bandwidth is determined by the satellite geometric mapping relationship and the filter gain. It is a time-varying parameter that changes dynamically with the observation scene. The theoretical equivalent noise bandwidth of each channel is

[0046] The channel-by-channel ENB sequence is retained, and the average is taken over the window.

[0047] (2) Predictive constraint early / late mismatch (NF2): If the first... Each channel in the historical era The prior code phase error is The local effective correlation amplitude is estimated as follows: Then the predicted Early and Late amplitudes can be approximately expressed as:

[0048]

[0049] when When within the discriminator's linear operating range, it can be respectively in If we perform a first-order Taylor expansion, the predicted early / late difference is:

[0050] in For the code autocorrelation function in The first derivative at that point. Correspondingly, the measured Early-Late difference is given by the current correlator and denoted as... Therefore, the mismatch statistic NF2 is defined as follows:

[0051] (3) Normalized State Correction Energy (NF3): During the Kalman filter measurement update process, the state correction vector is calculated from the filter gain and the innovation vector. The state correction vector caused by the measurement update is...

[0052] Utilizing prior covariance Normalizing this correction value, the normalized state correction energy is: The corresponding window statistics are

[0053] When the cleverness bias has not yet been fully absorbed by the covariance adaptation The response will increase significantly; if the adaptive noise model is too strong, the response of this feature may be partially suppressed.

[0054] (4) Normalized innovation square (NF4): The normalized innovation square is defined as... The final window statistics are taken as

[0055] In the absence of dexterity Approximately obey Distribution. And compared to, It more directly reflects whether the measurement residuals exceed the range allowed by the theoretical covariance.

[0056] (5) Whitening news time series statistics (NF5): To measure the temporal correlation of news series, we first define the scalarized whitening news series as...

[0057] in The measurement vector for the current epoch. dimensionality for A vector of all 1s. For any window Let the sample mean be denoted as Then the first The sample autocorrelation function of order 1 can be expressed as:

[0058] Therefore, the first One channel in the window The Ljung-Box statistic on the top is

[0059] Finally, NF5 takes the average of the statistics for each channel, i.e.

[0060] It preserves the temporal correlation between each satellite channel, avoiding the loss of statistical information after compressing the information from different channels into a single scalar.

[0061] (6) Code-carrier distance rate consistency residual (NF6): Theoretically, the code delay change rate and carrier Doppler both correspond to the same radial velocity, therefore they should remain consistent after being converted to the distance rate dimension. Let... For the first Estimation of the rate of change of code delay for each channel. For the corresponding carrier Doppler estimation, Given the L1 carrier wavelength, the code-carrier distance rate consistency residual is:

[0062] If there is a slight mismatch between the smart jamming and the code delay and Doppler joint forgery in a single channel, This will increase accordingly. The final window statistic is defined as follows.

[0063] (7) Residual Subspace Mode Dominance Ratio (NF7): If all epoch-level satellite geometric matrices are available, the orthogonal complementary projection matrix of the measurement subspace can be explicitly constructed to analyze the residual subspace structure. Let... Let be the projection matrix on the orthogonal complement of the measurement subspace, where for An identity matrix of dimension 1. Projecting the innovation vector onto the residual subspace yields:

[0064] In the window Construct the covariance matrix of the residual subspace above, i.e.

[0065] Therefore, the mode dominance ratio of the residual subspace is defined as...

[0066] in This represents the largest eigenvalue. Without dexterity, the energy distribution in the residual subspace is more isotropic; a single dexterity source injects a consistent structural bias into the residual subspace, thus significantly increasing the proportion of the largest eigenvalue. Therefore, It can effectively describe the low-rank aggregation phenomenon under single-source smart attacks.

[0067] (8) Cross-channel correlation mean of whitening information (NF8): NF8 and NF5 share the same whitening information sequence. For the window , define the first , The sample mean of the two channels is

[0068] Therefore, its Pearson correlation coefficient can be expressed as:

[0069] Finally, the cross-channel average of the absolute correlation coefficient was used as the global feature:

[0070] It forms a complementary effect with NF5: NF5 is used to describe the temporal correlation of a single channel, while NF8 is used to characterize the spatial correlation between different channels. Together, they constrain the statistical hypothesis that the innovations should be approximately independent after whitening.

[0071] (9) Dominance of the de-geometric Doppler common mode (NF9): Let For the measured Doppler vector, For the prior state The predicted Doppler vector is then defined as the geometric Doppler residual.

[0072] To suppress the common clock drift term in the receiver, a mean-free residual vector is further constructed, which yields the following result:

[0073] Then at the window Remove the window mean of each channel from the inner layer, i.e. Based on this, the window covariance matrix is ​​constructed as follows:

[0074] Therefore, the dominance of the geometric Doppler common mode is defined as...

[0075] In real-world scenarios, the Doppler residuals of each satellite are primarily determined by the differences in independent geometric projections. Doppler perturbations under single-source agility often share the same clock source and the same motion trajectory, thus exhibiting significant common modes. Therefore, NF9[k] can effectively describe the low-rank common-mode structure of single-source agility in the Doppler domain.

[0076] The baseband common-mode fingerprint layer feature design includes: For digital intermediate frequency injection-type smart devices, all smart signals are generated by the same smart transmitter and the same digital-to-analog link. The smart signals inevitably share the same common-mode disturbance source, i.e., the RF fingerprint feature.

[0077] (1) Noise Baseline Dominance Ratio (BBRF1): BBRF1 is used to describe whether there is abnormal synchronization in the noise baseline power fluctuations between different satellite channels. Each channel in the historical era The Early, Late, and Prompt power values ​​are respectively , , The noise basis estimator is defined as

[0078] For window Construct a centered noise matrix, i.e.

[0079] in Therefore, the covariance matrix is

[0080] The final noise basis dominant mode ratio is defined as follows:

[0081] In the absence of single-source common-mode disturbances, its null hypothesis baseline is approximately: The noise floor fluctuations are more independent across channels, and the covariance matrix does not exhibit a significant dominant mode; conversely, if multiple channels are disturbed by the same attack source, the proportion of the largest eigenvalue will increase significantly.

[0082] (2) Cross-satellite amplitude co-fluctuation correlation mean (BBRF2): BBRF2 describes whether there are abnormal synchronous fluctuations in the amplitudes of different satellites' Prompt values, in order to detect abnormal cross-satellite coupling structures. The Prompt amplitude of each channel is First, perform relative normalization within the window, that is...

[0083]

[0084] Therefore, the channel pair The correlation coefficient of amplitude co-fluctuation can be expressed as

[0085] The global mean feature is constructed as

[0086] For real satellite signals, amplitude fluctuations under different propagation paths are approximately independent; for single-source smart signals, all smart signals share the same gain ripple and the same clock disturbance, therefore It will increase significantly.

[0087] (3) Differential Code-Doppler Consistency Divergence (BBRF3): NF6 represents the single-channel absolute residual, while BBRF3 is the overall divergence for further detecting multi-channel differential residuals, and Let the first channel be the reference channel, and the difference residual vector can be constructed as follows:

[0088] Therefore, all common items are eliminated, leaving only relative inconsistencies across channels. For windows... Let its sample mean be denoted as Then the sample covariance is

[0089] And by constructing a Hotelling-type statistic, the differential code-Doppler uniformity divergence is defined as follows:

[0090] If the smart device cannot maintain a strictly consistent code delay-Doppler consistency relationship simultaneously on all satellite channels, then The systematic deviation from zero will significantly increase BBRF3; if the residuals contain only common disturbances from the receiver, the differential step will cancel them out first, and the statistic will not increase.

[0091] In one embodiment, the baseband correlation layer features include signal power correlation indicators, correlation peak distortion indicators, and amplitude ratio indicators of the main peak and secondary peaks.

[0092] In one embodiment, the navigation filter layer features characterizing the closed-loop operation state of the extended Kalman filter module based on the vector tracking loop are extracted from the output data, including: calculating the equivalent noise bandwidth of the vector delay-locked loop for each channel based on the geometric mapping matrix and the filter gain matrix of the vector tracking loop; calculating the prediction constraint early-late mismatch degree based on the code phase deviation obtained from the filter prediction and the measured early-late branch amplitude; performing multi-dimensional statistics on amplitude, temporal correlation, and subspace distribution on the innovation sequence and residual sequence to obtain filter-type features characterizing the abnormal operation of the closed-loop loop; and obtaining the navigation filter layer features based on the equivalent noise bandwidth of the vector delay-locked loop, the prediction constraint early-late mismatch degree, and the filter-type features.

[0093] In one embodiment, calculating the equivalent noise bandwidth of the vector delay-locked loop for each channel based on the geometric mapping matrix and filter gain matrix of the vector tracking loop includes: acquiring the geometric mapping matrix and extended Kalman filter position-clock difference gain sub-matrix of the vector tracking loop epoch-by-epoch; solving for the equivalent noise bandwidth of the vector delay-locked loop for each channel epoch-by-epoch based on the geometric mapping matrix and the extended Kalman filter position-clock difference gain sub-matrix; and calculating the mean of the equivalent noise bandwidth of a single channel epoch-by-epoch according to a fixed-length non-overlapping time window to obtain the equivalent noise bandwidth characteristics of the vector delay-locked loop. The equivalent noise bandwidth characteristics of the vector delay-locked loop are as follows:

[0094] in, The equivalent noise bandwidth characteristics of the vector delay-locked loop. For the first The first channel The equivalent noise bandwidth of the vector delay-locked loop in an epoch. The number of sampling points within a single time window. For the first The set of epochs corresponding to each window.

[0095] In one embodiment, calculating the prediction-constrained early-late mismatch degree based on the code phase deviation obtained from the filter prediction and the measured early-late branch amplitudes includes: obtaining the code phase deviation from the extended Kalman filter prediction result, calculating the predicted early-late branch amplitudes based on the code phase deviation and the code autocorrelation function; reading the measured early-late branch amplitudes output by the multi-channel correlator, calculating the difference between the predicted early-late branch amplitudes and the measured early-late branch amplitudes; calculating the difference between the predicted early-late branch amplitudes and the measured early-late branch amplitudes, normalizing the difference based on the sum of the amplitudes of the early and late branches, and then performing statistical analysis through a fixed-length non-overlapping time window to obtain the prediction-constrained early-late mismatch degree; the prediction-constrained early-late mismatch degree is:

[0096] in, To predict the early or late mismatch of constraints, For the first The first channel Measured difference in time of epoch, For the first The first channel Epoch-based filtered prediction of early / late difference, The signal amplitude of the early branch. The signal amplitude of the delayed branch, The number of sampling points within a single time window. For the first The set of epochs corresponding to each window.

[0097] In one embodiment, performing multi-dimensional statistics on amplitude, temporal correlation, and subspace distribution on the innovation sequence and residual sequence to obtain filter-type features characterizing loop closed-loop operation anomalies includes: calculating the normalized state correction energy and the normalized innovation square to obtain innovation amplitude dimension statistics; calculating the whitened innovation temporal deviation statistics to obtain innovation temporal correlation dimension statistics; calculating the code-carrier distance rate consistency residual, residual subspace mode dominance ratio, whitened innovation cross-channel correlation mean, and de-geometric Doppler common mode dominance ratio to obtain residual and cross-channel correlation dimension statistics; and obtaining filter-type features characterizing loop closed-loop operation anomalies based on the innovation amplitude dimension statistics, innovation temporal correlation dimension statistics, and residual and cross-channel correlation dimension statistics.

[0098] Specifically, the normalized state correction energy characteristic is obtained based on the filter state correction quantity and the prior error covariance matrix. The normalized innovation squared characteristic is obtained based on the innovation vector and the innovation covariance matrix. The whitening innovation time-series deviation statistic characteristic is obtained based on the multi-order autocorrelation coefficients of the whitening innovation sequence. The code-carrier distance rate consistency residual characteristic is obtained based on the physical correspondence between the code delay change rate and carrier Doppler. The residual subspace mode dominance ratio characteristic is obtained based on the eigenvalue distribution of the residual subspace covariance matrix. The cross-channel correlation mean characteristic of the whitening innovation is obtained based on the cross-correlation coefficients of the whitening innovations in each channel. The de-geometric Doppler common mode dominance ratio characteristic is obtained based on the covariance distribution of the de-geometric Doppler residuals.

[0099] In one embodiment, the common-mode fingerprint layer features include the proportion of main modes in a multi-channel noise substrate, the correlation features of signal amplitude linkage across all satellite channels, and the cross-channel code-Doppler consistency divergence features.

[0100] In one embodiment, the method further includes: performing spatial coupling modeling based on multi-layer physical features to output a spatial coupling feature representation suitable for GNSS smart detection; performing spatial coupling modeling based on multi-layer physical features includes: dividing the multi-layer physical features into node features, global features, and edge features; constructing a shared physical prior graph based on the multi-satellite spatial coupling prior of the vector tracking loop, and constructing an anomaly response graph for the current epoch based on the edge features; adaptively fusing the shared physical prior graph and the anomaly response graph based on the physical anomaly intensity to generate a dynamic adjacency graph; performing graph convolutional encoding on the node features based on the dynamic adjacency graph, and fusing the global features to obtain the spatial coupling representation corresponding to each epoch.

[0101] In this embodiment, the multi-satellite channels of the vector tracking receiver exhibit inherent spatial coupling relationships due to geometric distribution, shared clock, and unified filtering. The original, flat, multi-layered physical features cannot structurally characterize the inter-channel correlation constraints, nor can they fully utilize multi-channel physical priors to enhance anomaly representation capabilities. Therefore, spatial coupling modeling is introduced to deeply integrate the physical prior structure with real-time anomaly features, further strengthening the feature's ability to characterize cross-channel linked anomalies. The output spatial coupling feature representation can be directly input into traditional classifiers such as Support Vector Machines and Random Forests, or into time-series detection models such as LSTM and Transformer. This helps downstream models better capture multi-channel collaborative anomalies, effectively improving detection accuracy and reducing false alarm probability in concealed and agile scenarios. In addition to spatial coupling representation, the generated multi-layered physical features can also be directly used in various scenarios such as traditional threshold detection, receiver operating status monitoring, interference type identification, and loop performance evaluation, demonstrating strong versatility and adaptability.

[0102] Specifically, two complementary graph structures are constructed at each epoch: a time-shared physical prior graph to describe stable coupling relationships, and an anomaly response graph to characterize the current agile-driven cross-channel anomaly correlations. Subsequently, the two graphs are adaptively fused using a physical gating mechanism to generate the dynamic topology required for graph convolution. Before graph convolution, this invention first standardizes 15 physical features based on normal sample statistics and classifies them into node features, edge features, and global features according to their spatial affiliation. Based on this, physical gating vectors and event scoring vectors are further constructed to form the structured input required for subsequent dual-source dynamic graph encoding and sparse temporal modeling of events.

[0103] (1) Node feature tensor: for an epoch The For each satellite, six features that can be calculated on a satellite-by-satellite basis and have a clearly defined channel affiliation are selected to form the node feature vector, i.e.

[0104] Stacking all satellite node features row-wise yields an epoch-level node feature matrix:

[0105] (2) Global Feature Vector: For features that are difficult to classify into a single satellite channel but can reflect the overall state of the filter or the common-mode structure of multiple channels, they are organized into a global feature vector:

[0106] (3) Edge feature vector: Edge features are used to describe the correlation between satellite node pairs. Let the set of satellite node pairs be...

[0107] In the experiment, the edge features were mainly provided by BBRF2 features. For any node... Define its edge characteristics as From this, we can obtain the epoch. The edge feature vector is

[0108] In a graph encoder, The upper triangular index is mapped to a symmetric edge weight matrix, which is used to construct the anomaly response graph and perform edge enhancement.

[0109] (4) Physical gating vector and event scoring vector: To enhance the correspondence between model input and agile mechanism, two low-dimensional physical driving vectors are further constructed before the graph encoder and the temporal encoder. Then, the star-by-star features NF1 and BB3 are aggregated by mask mean on the effective channels, i.e.

[0110] Then, a graph fusion gate vector is constructed:

[0111] It covers key links in the smart interference propagation chain and can serve as a physical prior input for subsequent dual-graph fusion gating. Specifically: (1) Shared physical prior graph: In order to characterize the relatively stable inter-satellite coupling prior within a short observation window, the model introduces a learnable adjacency parameter matrix shared across time. To satisfy the symmetry constraint of the adjacency matrix of an undirected graph, the shared physical prior graph is defined as follows:

[0112] in This is the Sigmoid function. Since this parameter matrix is ​​shared across all epochs, It can learn long-term stable multi-star space coupling modes. (2) Anomaly response graph: To characterize the cross-channel common mode anomaly in the current epoch, the edge feature vectors are rearranged into a symmetric matrix. ,in Indicates the first With the Each satellite channel in the historical period The amplitude co-oscillation correlation intensity. Subsequently, the anomalous response map is obtained by linear scaling and Sigmoid mapping, i.e.

[0113] in and These are learnable scalar parameters.

[0114] (3) Since the anomaly response graph has been explicitly constructed from edge features, edge enhancement is only performed once on the physical prior graph to supplement the anomaly association information of the current epoch, i.e.

[0115] in For Hadamard products, It is a matrix of all 1s. This is a learnable scalar. Subsequently, the model utilizes physical gating vectors. Generate dual-image fusion coefficients:

[0116] And complete the adaptive fusion of the physical prior map and the anomalous response map:

[0117] Therefore, when the receiver is in a stable tracking phase, the fused graph tends to be more based on the shared physical prior graph; when smart takeover leads to cross-channel anomalous enhancement, the fused graph gradually shifts to the dynamic anomalous response graph.

[0118] (4) Residual graph convolution and global readout: Before graph convolution, the model first supplements the effective nodes with self-loops and uses symmetric normalization to improve propagation stability, i.e.

[0119] in, This is the degree matrix. Then, two layers of graph convolution are performed, i.e.

[0120]

[0121] To alleviate the oversmoothing and gradient decay issues in deep graph convolution, the model further introduces residual coupling. Assume linear projection... Change the node feature dimension from Mapped to Then the node representation after residual normalization is:

[0122] remember for The Middle The model embeds nodes. To obtain an epoch-level graph representation, it employs three readout methods simultaneously: masked attention pooling, masked max pooling, and global feature projection. The masked attention weights are defined as follows:

[0123] From this, we can derive three types of readout representations, namely...

[0124] In this process, attention pooling highlights key anomalies, max pooling retains the strongest local responses, and global feature projection supplements the system-level statistical information of the navigation filter layer and the common-mode fingerprint layer. Finally, the three representations are concatenated and input into the multilayer perceptron to obtain the epoch-level spatially coupled representation: .

[0125] In one embodiment, classifying multi-layer physical features into node features, global features, and edge features includes: classifying baseband correlation layer features and navigation filter layer features that can be independently calculated for each satellite channel as node features; classifying global statistical features that reflect the overall state of the receiver system and the cross-channel common-mode structure as global features; and classifying common-mode statistical features that characterize the correlation strength between satellite channel pairs as edge features.

[0126] 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.

[0127] In one embodiment, a multi-layer physical feature extraction device for a GNSS vector tracking receiver is provided, comprising: The signal acquisition module is used to acquire the received signal from the GNSS vector tracking receiver; The baseband correlation layer feature extraction module is used to perform baseband correlation processing on the received signal to obtain multi-channel correlator output data, and extract baseband correlation layer features that characterize the correlation peak shape based on the multi-channel correlator output data; The navigation filter layer feature extraction module is used to extract navigation filter layer features that characterize the closed-loop operation state of the extended Kalman filter module based on the vector tracking loop. The output data includes navigation state estimation, innovation residuals, and feedback control data. The common-mode fingerprint layer feature extraction module is used to extract common-mode fingerprint layer features that characterize cross-channel linkage based on the joint statistical results of all satellite channels; The multi-layer physical feature output module is used to perform statistical quantization and standardization on the baseband correlation layer features, navigation filter layer features, and common mode fingerprint layer features using a fixed-length non-overlapping window to obtain the multi-layer physical features corresponding to each epoch.

[0128] Specific limitations regarding the multi-layer physical feature extraction device for GNSS vector tracking receivers can be found in the limitations of the multi-layer physical feature extraction method for GNSS vector tracking receivers mentioned above, and will not be repeated here. Each module in the aforementioned multi-layer physical feature extraction device for GNSS vector tracking receivers can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the operations corresponding to each module.

[0129] 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.

[0130] 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 the invention. 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 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 multi-layer physical features from a GNSS vector tracking receiver, characterized in that, The method includes: Acquire the received signal from the GNSS vector tracking receiver; Baseband correlation processing is performed on the received signal to obtain multi-channel correlator output data, and baseband correlation layer features characterizing the correlation peak morphology are extracted based on the multi-channel correlator output data. The output data of the extended Kalman filter module based on the vector tracking loop is used to extract navigation filter layer features that characterize the closed-loop operation state of the loop; the output data includes navigation state estimation, innovation residuals and feedback control data; Common-mode fingerprint layer features characterizing cross-channel linkage are extracted based on joint statistical results from all satellite channels; The features of the baseband correlation layer, navigation filter layer, and common mode fingerprint layer are statistically quantized and standardized using a fixed-length non-overlapping window to obtain the multi-layer physical features corresponding to each epoch.

2. The method according to claim 1, characterized in that, The baseband correlation layer features include signal power correlation indicators, correlation peak distortion indicators, and amplitude ratio indicators of the main peak and secondary peaks.

3. The method according to claim 1, characterized in that, The output data extraction of the extended Kalman filter module based on the vector tracking loop, which characterizes the navigation filter layer features representing the closed-loop operation state of the loop, includes: Based on the geometric mapping matrix and filter gain matrix of the vector tracking loop, the equivalent noise bandwidth of the vector delay-locked loop for each channel is calculated. The prediction constraint early-late mismatch degree is calculated based on the code phase deviation obtained from the filter prediction and the measured early-late branch amplitude. Multi-dimensional statistics on amplitude, temporal correlation, and subspace distribution are performed on the innovation sequence and residual sequence to obtain filter characteristics that characterize loop closed-loop operation anomalies; The navigation filter layer features are obtained based on the equivalent noise bandwidth of the vector delay-locked loop, the prediction constraint early-late mismatch, and the filter class features.

4. The method according to claim 3, characterized in that, The calculation of the equivalent noise bandwidth of the vector delay-locked loop for each channel based on the geometric mapping matrix and the filter gain matrix of the vector tracking loop includes: The geometric mapping matrix and extended Kalman filter position-clock gain submatrix of the vector tracking loop are acquired epoch by epoch. The equivalent noise bandwidth of the vector delay-locked loop for each channel is solved based on the geometric mapping matrix and the extended Kalman filter position-clock gain submatrix. The equivalent noise bandwidth of a single channel is averaged over a fixed-length non-overlapping time window to obtain the equivalent noise bandwidth characteristics of the vector delay-locked loop; the equivalent noise bandwidth characteristics of the vector delay-locked loop are as follows: in, The equivalent noise bandwidth characteristics of the vector delay-locked loop. For the first The first channel The equivalent noise bandwidth of the vector delay-locked loop in an epoch. The number of sampling points within a single time window. For the first The set of epochs corresponding to each window.

5. The method according to claim 3, characterized in that, The calculation of the prediction constraint early-late mismatch degree based on the code phase deviation obtained from the filter prediction and the measured early-late branch amplitude includes: The code phase deviation is obtained from the prediction results of the extended Kalman filter, and the predicted early and late branch amplitudes are calculated based on the code phase deviation and the code autocorrelation function. Read the measured early and late branch amplitudes output by the multi-channel correlator, and calculate the difference between the predicted early and late branch amplitudes and the measured early and late branch amplitudes. The difference between the predicted early and late branch amplitudes and the measured early and late branch amplitudes is calculated. This difference is normalized based on the sum of the amplitudes of the early and late branches. Then, statistics are performed using a fixed-length non-overlapping time window to obtain the prediction constraint early-late mismatch degree. The prediction constraint early-late mismatch degree is: in, To predict the early or late mismatch of constraints, For the first The first channel Measured difference in time of epoch, For the first The first channel Epoch-based filtered prediction of early / late difference, The signal amplitude of the early branch. The signal amplitude of the delayed branch, The number of sampling points within a single time window. For the first The set of epochs corresponding to each window.

6. The method according to claim 3, characterized in that, The multi-dimensional statistical analysis of amplitude, temporal correlation, and subspace distribution of the innovation sequence and residual sequence yields filter-type features characterizing loop closed-loop operation anomalies, including: Calculate the normalized state correction energy and the normalized squared innovation to obtain statistical results of the innovation amplitude dimension. Calculate the whitening innovation time series deviation statistic to obtain the statistical results of the innovation time series correlation dimension; Calculate the code-carrier distance rate consistency residual, residual subspace mode dominance ratio, whitening innovation cross-channel correlation mean, and de-geometric Doppler common mode dominance ratio to obtain statistical results of residual and cross-channel correlation dimensions. Based on the statistical results of the innovation amplitude dimension, the statistical results of the innovation time series correlation dimension, and the statistical results of the residual and cross-channel correlation dimension, filter-type features that characterize the abnormal operation of the loop closed loop are obtained.

7. The method according to claim 1, characterized in that, The common-mode fingerprint layer features include the proportion of main modes in the multi-channel noise substrate, the correlation features of the signal amplitude linkage across all satellite channels, and the cross-channel code-Doppler consistency divergence features.

8. The method according to claim 1, characterized in that, The method further includes: Spatial coupling modeling is performed based on the aforementioned multi-layer physical characteristics, outputting spatial coupling feature representations suitable for GNSS smart detection; spatial coupling modeling based on multi-layer physical characteristics includes: The multi-layer physical features are divided into node features, global features, and edge features; A shared physical prior graph is constructed based on the multi-star spatial coupling prior of the vector tracking loop, and an anomaly response graph for the current epoch is constructed based on the edge features. A dynamic adjacency graph is generated by adaptively fusing the shared physical prior graph and the anomaly response graph based on the intensity of physical anomalies. The node features are subjected to graph convolutional encoding based on the dynamic adjacency graph, and the global features are fused to obtain the spatial coupling representation corresponding to each epoch.

9. The method according to claim 1, characterized in that, The multi-layer physical features are divided into node features, global features, and edge features, including: The baseband correlation layer features and navigation filter layer features, which can be calculated independently for each satellite channel, are divided into node features; Global statistical features that reflect the overall state of the receiver system and the cross-channel common-mode structure are classified as global features. The common-mode statistical features that characterize the correlation strength between satellite channel pairs are divided into edge features.

10. A multi-layer physical feature extraction device for a GNSS vector tracking receiver, characterized in that, The device includes: The signal acquisition module is used to acquire the received signal from the GNSS vector tracking receiver; The baseband correlation layer feature extraction module is used to perform baseband correlation processing on the received signal to obtain multi-channel correlator output data, and extract baseband correlation layer features characterizing the correlation peak shape based on the multi-channel correlator output data; The navigation filter layer feature extraction module is used to extract navigation filter layer features that characterize the closed-loop operation state of the extended Kalman filter module based on the output data of the vector tracking loop; the output data includes navigation state estimation, innovation residual and feedback control quantity data; The common-mode fingerprint layer feature extraction module is used to extract common-mode fingerprint layer features that characterize cross-channel linkage based on the joint statistical results of all satellite channels; The multi-layer physical feature output module is used to perform statistical quantization and standardization on the baseband correlation layer features, navigation filter layer features, and common mode fingerprint layer features using a fixed-length non-overlapping window to obtain the multi-layer physical features corresponding to each epoch.

Images