Unmanned aerial vehicle borne gnss jammer robust positioning method based on physical guidance learning

Abstract

The application discloses a UAV-borne GNSS interference source robust positioning method based on physical guidance learning, relates to the field of communication technology, and determines multi-domain signal representation and physical model angle priori according to array I / Q snapshot data, calculates reliable azimuth constraints based on a target double-flow neural network; calculates the initial position of the interference source according to a plurality of reliable azimuth constraints; processes the initial position of the interference source through adaptive exponential moving average filtering to obtain a smooth position estimate, and determines the spatial geometric constraint and weighted geometric residual corresponding to each sensing moment; determines the target position of the interference source according to the smooth position estimate, the received signal effectiveness index, the spatial geometric constraint and the weighted geometric residual corresponding to each sensing moment. Through the above technical scheme, the influence of complex propagation factors such as reflection, diffraction, scattering, shielding and noise on the angle of arrival estimation can be reduced, the accuracy and robustness of the angle of arrival estimation are improved, and the positioning accuracy of the target position of the interference source is improved.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and specifically to a robust localization method for UAV-borne GNSS interference sources based on physics-guided learning. Background Technology

[0002] The Global Navigation Satellite System (GNSS) provides fundamental positioning services for aviation, maritime, transportation, power, communications, and emergency rescue. With the emergence of low-cost jamming devices, GNSS interference incidents are increasing. Interference signals can suppress or disrupt the receiver's normal acquisition and tracking of satellite signals, leading to decreased positioning accuracy, positioning interruptions, and even navigation service failure. Therefore, quickly and accurately determining the location of GNSS interference sources is crucial for on-site investigation, interference mitigation, and positioning service restoration. However, traditional ground-based inspection methods typically rely on manually carried or vehicle-mounted spectrum monitoring equipment, which is easily limited by road conditions and building obstructions, resulting in low positioning efficiency. Unmanned aerial vehicles (UAVs) offer advantages such as high mobility and flexible observation perspectives, allowing them to receive interference signals from different spatial locations. Therefore, research on UAV-based GNSS interference source localization has significant practical implications.

[0003] Existing UAV interference source localization methods mostly rely on geometric calculations based on observations such as received signal strength, angle of arrival, time difference of arrival, or frequency difference of arrival. Among them, methods based on received signal strength usually rely on path loss models and are susceptible to unknown transmit power, blockage, and multipath fading; methods based on angle of arrival need to extract stable angle information from the array received signal, but in complex propagation environments, reflection, diffraction, scattering, and non-line-of-sight propagation can cause the main transmission direction to deviate from the actual interference source direction; methods based on time difference or frequency difference have high requirements for synchronization accuracy, motion state, and propagation path stability, making them difficult to implement in denial scenarios.

[0004] In reality, buildings, ground, vehicles, and other obstacles in the environment cause GNSS interference signals to be absorbed, reflected, diffracted, and scattered during propagation, resulting in complex multipath superposition characteristics of the receiver array signal. Traditional methods based on ideal free-space propagation or fixed analytical models are insufficient to accurately characterize these complex electromagnetic propagation processes. Summary of the Invention

[0005] The purpose of this invention is to provide a robust localization method for UAV-borne GNSS interference sources based on physics-guided learning, in order to solve the technical problems existing in related technologies.

[0006] To achieve the above objectives, this invention provides a robust localization method for UAV-borne GNSS interference sources based on physics-guided learning, comprising:

[0007] The array I / Q snapshot data of the UAV at each sensing moment and the feature information of the UAV at each sensing moment are acquired. The feature information includes the target position, attitude information and received signal validity index of the UAV. The UAV is equipped with an antenna array.

[0008] For each sensing moment, the array I / Q snapshot data and feature information are used to determine the multi-domain signal representation and physical model angle prior based on the array I / Q snapshot data. Based on the target dual-stream neural network, the reliable orientation constraint corresponding to each sensing moment is calculated according to the multi-domain signal representation, physical model angle prior and feature information. After processing all array I / Q snapshot data and feature information corresponding to each sensing moment, multiple reliable azimuth constraints are obtained, and the initial position of the interference source corresponding to each sensing moment is calculated based on the multiple reliable azimuth constraints. For the initial position of the interference source corresponding to each sensing time, the following processing is performed: the initial position of the interference source is processed by adaptive exponential moving average filtering to obtain a smooth position estimate, and the spatial geometric constraints and weighted geometric residuals corresponding to each sensing time are determined; The target location of the interference source is determined based on the smoothed position estimate corresponding to each sensing moment, the received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals.

[0009] Optionally, the step of determining the multi-domain signal representation and physical model angle prior based on array I / Q snapshot data, and calculating the reliable azimuth constraint corresponding to each sensing moment based on the multi-domain signal representation, physical model angle prior, and feature information using a target two-stream neural network, includes: Extract the covariance matrix from the array I / Q snapshot data, construct a multi-domain signal representation based on the covariance matrix, and construct a physical model angle prior based on the covariance matrix through electromagnetic propagation mechanism and array receiving model; The target dual-stream neural network is used to process the multi-domain signal representation and the physical model angle prior to obtain the arrival angle probability distribution, confidence level and non-line-of-sight probability. The arrival angle estimate is calculated based on the physical model angle prior and the arrival angle probability distribution. The target position, arrival angle estimate, arrival angle probability distribution, confidence level and non-line-of-sight probability are fused into a reliability-aware angle observation. The reliability weights are calculated based on the reliability perception angle sensing, target position, attitude information, confidence level, non-line-of-sight probability, and received signal effectiveness index. The reliable orientation constraints of the UAV under test are then constructed based on the target position, reliability weights, and angle of arrival estimates.

[0010] Optionally, the step of calculating the initial position of the interference source corresponding to each sensing moment based on multiple reliable orientation constraints includes: A weighted geometric residual model is constructed based on multiple reliable orientation constraints. The weighted geometric residual model is then solved iteratively using the iterative reweighted least squares method to obtain the fusion weights of the UAV under test at each sensing time. The initial position of the interference source at each sensing time is then calculated based on the fusion weights.

[0011] Optionally, determining the target location of the interference source based on the smoothed position estimate, received signal effectiveness index, spatial geometric constraints, and weighted geometric residual corresponding to each sensing time includes:

[0012] Multiple sensing moments are sorted in ascending order, and the smoothed position estimate, received signal effectiveness index, spatial geometric constraint, and weighted geometric residual corresponding to each sensing moment are compared with preset conditions. When the smoothed position estimate, received signal effectiveness index, spatial geometric constraint, and weighted geometric residual satisfy the preset conditions, the smoothed position estimate, received signal effectiveness index, spatial geometric constraint, and weighted geometric residual corresponding to each sensing moment in the target sensing moment are determined. The preset conditions are: the spatial geometric constraint is greater than or equal to the spatial geometric constraint strength threshold; the absolute value of the difference between the smoothed position estimate and the smoothed position estimate of the previous sensing moment is less than or equal to the step change threshold; the difference between the smoothed position estimate corresponding to the first sensing moment and the smoothed position estimate corresponding to the last sensing moment within the preset time window is less than or equal to the window drift threshold; the received signal effectiveness index is greater than or equal to the received signal effectiveness threshold; and the difference between the weighted geometric residual and the smoothed position estimate is less than or equal to the preset consistency threshold. The target sensing moment is the current sensing moment and all previous sensing moments.

[0013] For each sensing moment in the target sensing time, a comprehensive quality score is calculated based on the smooth position estimation, received signal effectiveness index, spatial geometric constraints, and weighted geometric residual corresponding to each sensing moment to obtain the initial position of the interference source. After the target sensing time is completed, multiple comprehensive quality scores are obtained.

[0014] Among multiple comprehensive quality scores, the initial position of the interference source corresponding to the highest comprehensive quality score is determined as the target position.

[0015] Optionally, the positioning method further includes:

[0016] When the smoothed position estimation, received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals do not meet the preset conditions, the estimated angle of arrival and initial position corresponding to the sensing moment are obtained, and the position of the sensing waypoint corresponding to the next sensing moment is updated based on the estimated angle of arrival and initial position.

[0017] Optionally, the step of extracting the covariance matrix from the array I / Q snapshot data and constructing a multi-domain signal representation based on the covariance matrix includes:

[0018] Calculate the covariance matrix of the array I / Q snapshot data;

[0019] The covariance matrix is ​​decomposed into real part information, imaginary part information, amplitude information and phase information, and the real part information, imaginary part information, amplitude information and phase information are determined as multi-domain signal representation.

[0020] Optionally, the covariance matrix is ​​expressed by the following formula:

[0021] ;

[0022] in, Let covariance matrix be the variance matrix. For array I / Q snapshot data, This is the conjugate transpose of the array I / Q snapshot data. This represents the number of time sampling points for array I / Q snapshot data;

[0023] The prior angle of the physical model is expressed by the following calculation formula:

[0024] ;

[0025] in, As a priori from the perspective of the physical model, This is the conjugate transpose of the array guide vector. The noise subspace is obtained by eigenvalue decomposition of the covariance matrix. This is the conjugate transpose of the noise subspace obtained by eigenvalue decomposition of the covariance matrix. The array guide vector for the candidate direction.

[0026] Optionally, the target dual-stream neural network includes a signal representation stream network and a physical prior stream network. The process of processing the multi-domain signal representation and physical model angle prior through the target dual-stream neural network to obtain the arrival angle probability distribution, confidence level, and non-line-of-sight probability includes:

[0027] The signal feature information in the multi-domain signal representation is extracted by the signal representation flow network, and the angle structure features in the angle prior of the physical model are extracted by the physical prior flow network.

[0028] The signal feature information and angular structure features are processed to obtain the arrival angle probability distribution, confidence level, and non-line-of-sight probability.

[0029] Optionally, calculating the estimated angle of arrival based on the prior angle of the physical model and the probability distribution of the angle of arrival includes:

[0030] The physical model angle prior is normalized to obtain the normalized physical model angle prior;

[0031] The posterior angle distribution is calculated based on the prior angle and the probability distribution of the angle of arrival from the normalized physical model.

[0032] Extract the estimated angle of arrival from the posterior distribution of the angle.

[0033] Optionally, the posterior angle distribution is expressed by the following formula:

[0034] ;

[0035] ;

[0036] ;

[0037] in, For the posterior distribution of angle, For the probability distribution of the angle of arrival, For adaptive fusion weights, For the normalized physical model perspective prior, For the first k The physical model spatial spectrum after normalization or centering of the candidate arrival angles For the first i The physical model spatial spectrum after normalization or centering of the candidate arrival angles The sharpening coefficient is the a priori distribution. This represents the total number of candidate arrival angle directions. For the first One candidate arrival angle, , Candidate arrival angle direction, For the target two-stream neural network in the candidate angle of arrival direction The predicted angle value output above; For the target two-stream neural network in the candidate direction The predicted angle value output above.

[0038] Optionally, the step of calculating the reliability weight based on reliability-aware angle sensing, target position, attitude information, confidence level, non-line-of-sight probability, and received signal effectiveness index includes:

[0039] The direction vector of the UAV in the angle observation is calculated based on the reliability perception angle sensing, target position and attitude information, and the normal vector of the UAV in the angle observation is calculated based on the direction vector.

[0040] Calculate the orthogonal geometric residual based on the normal vector and the target position;

[0041] Reliability weights are calculated based on orthogonal geometric residuals, confidence levels, non-line-of-sight probability, and received signal effectiveness indices.

[0042] Optionally, the weighted geometric residual model is expressed by the following formula:

[0043] ;

[0044] in, For a weighted geometric residual model, This represents the total number of candidate arrival angle directions. The reliability weight in the reliable orientation constraint. Let be the two-dimensional horizontal position vector of the GNSS interference source to be solved. The target position in the reliable orientation constraint, This is the estimated angle of arrival in the reliable azimuth constraint. This is a vector transpose operation. An index for reliable orientation constraints within a time window. =1, 2, ..., ; It is a two-dimensional real number space;

[0045] The weighted geometric residual model is solved using a weighted method equation, which is expressed by the following formula:

[0046] ;

[0047] ;

[0048] ;

[0049] in, For spatial geometric constraint strength, For weighted observation vectors, For regularization terms, It is the normal vector. This is the conjugate transpose of the normal vector;

[0050] The fusion weight is expressed by the following formula:

[0051] ;

[0052] ;

[0053] in, For the first The fusion weights for each observation time. As a reliability weight, For robust weights, For the first The geometric residuals corresponding to each observation time. This is the residual tolerance parameter;

[0054] The smoothed position estimate is expressed by the following formula:

[0055] ;

[0056] in, To smooth the position estimation, For the first The smoothing coefficient at each observation time. For the first Smooth position estimation at each sensing moment For the first The initial location of the interference source at each observation time.

[0057] The above technical solution acquires array I / Q snapshot data and UAV feature information at each sensing moment. For each sensing moment, multi-domain signal representation and physical model angle prior are determined based on the array I / Q snapshot data. Based on the target dual-stream neural network, reliable azimuth constraints for each sensing moment are calculated according to the multi-domain signal representation, physical model angle prior, and feature information. Joint processing of the array I / Q snapshot data and the target dual-stream neural network allows the network to learn signal distortion characteristics in complex environments while being subject to physical prior constraints, thereby improving performance against reflection, diffraction, scattering, occlusion, and noise. The accuracy and robustness of the angle-of-arrival estimation are improved. After processing all array I / Q snapshot data and feature information corresponding to each sensing time, multiple reliable azimuth constraints are obtained. Based on these constraints, the initial position of the interference source at each sensing time is calculated. The initial position of the interference source is then processed using an adaptive exponential moving average filter to obtain a smoothed position estimate. The spatial geometric constraints and weighted geometric residuals for each sensing time are determined. The target position of the interference source is determined based on the smoothed position estimate, received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals for each sensing time, which improves the accuracy of the target position.

[0058] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0059] Figure 1 This is a schematic diagram illustrating a robust localization method for UAV-borne GNSS interference sources based on physical guided learning, according to an exemplary embodiment of the present invention.

[0060] Figure 2 This is a schematic diagram illustrating a scenario for locating GNSS interference sources on a UAV according to an exemplary embodiment of the present invention.

[0061] Figure 3This is a schematic diagram illustrating the structure of a physically guided target dual-stream neural network according to an exemplary embodiment of the present invention.

[0062] Figure 4 This is a schematic diagram illustrating the waypoint update process of a UAV according to an exemplary embodiment of the present invention. Detailed Implementation

[0063] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, so as to provide a better understanding of the concept of the present invention, the technical problem solved, the technical features constituting the technical solution, and the technical effects brought about.

[0064] like Figure 1 As shown, Figure 1 This is a schematic diagram illustrating a robust localization method for UAV-borne GNSS interference sources based on physical guided learning according to an exemplary embodiment of the present invention. (Refer to...) Figure 1 The method includes;

[0065] S101: Acquire array I / Q snapshot data of the UAV at each sensing moment and feature information of the UAV at each sensing moment, wherein the feature information includes the target position, attitude information and received signal validity index of the UAV, and the UAV is equipped with an antenna array;

[0066] S102: For the array I / Q snapshot data and feature information corresponding to each sensing moment, determine the multi-domain signal representation and physical model angle prior based on the array I / Q snapshot data, and calculate the reliable orientation constraint corresponding to each sensing moment based on the target dual-stream neural network, according to the multi-domain signal representation, physical model angle prior and feature information.

[0067] S103: After processing all array I / Q snapshot data and feature information corresponding to each sensing moment, multiple reliable azimuth constraints are obtained, and the initial position of the interference source corresponding to each sensing moment is calculated based on the multiple reliable azimuth constraints.

[0068] S104: For the initial position of the interference source corresponding to each sensing time, the following processing is performed: The initial position of the interference source is processed by adaptive exponential moving average filtering to obtain a smooth position estimate, and the spatial geometric constraints and weighted geometric residuals corresponding to each sensing time are determined.

[0069] S105: Determine the target location of the interference source based on the smoothed position estimate, received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals corresponding to each sensing moment.

[0070] The above technical solution acquires array I / Q snapshot data and UAV feature information at each sensing moment. For each sensing moment, multi-domain signal representation and physical model angle prior are determined based on the array I / Q snapshot data. Based on the target dual-stream neural network, reliable azimuth constraints for each sensing moment are calculated according to the multi-domain signal representation, physical model angle prior, and feature information. Joint processing of the array I / Q snapshot data and the target dual-stream neural network allows the network to learn signal distortion characteristics in complex environments while being subject to physical prior constraints, thereby improving performance against reflection, diffraction, scattering, occlusion, and noise. The accuracy and robustness of the angle-of-arrival estimation are improved. After processing all array I / Q snapshot data and feature information corresponding to each sensing time, multiple reliable azimuth constraints are obtained. Based on these constraints, the initial position of the interference source at each sensing time is calculated. The initial position of the interference source is then processed using an adaptive exponential moving average filter to obtain a smoothed position estimate. The spatial geometric constraints and weighted geometric residuals for each sensing time are determined. The target position of the interference source is determined based on the smoothed position estimate, received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals for each sensing time, which improves the accuracy of the target position.

[0071] To enable those skilled in the art to better understand the robust localization method for UAV-borne GNSS interference sources based on physics-guided learning provided by this invention, the above steps are illustrated in detail below.

[0072] For example, an antenna array can be an antenna system that combines multiple independent antenna elements in a specific geometric arrangement to collaboratively transmit and receive electromagnetic waves. In embodiments of the present invention, such as... Figure 2 As shown, an antenna array is mounted on the drone to receive signals emitted by interference sources; these signals can be used as array I / Q snapshot data. Figure 2 In this propagation environment, there are reflections, diffractions, scattering, absorption, noise, and direct paths in the array I / Q snapshot data. The array I / Q snapshot data can be a complex set of observation datasets obtained by a multi-antenna array receiver simultaneously acquiring I and Q orthogonal samples of all array elements at multiple times. This dataset serves as the raw, underlying input data for array signal processing, DOA direction finding, and target localization. The UAV's characteristic information includes target position, attitude information, and received signal effectiveness indicators. The target position can be the UAV's position coordinates at the current sensing time; the attitude information can be the UAV's angle or azimuth information at the current sensing time; and the received signal effectiveness indicator can be a quantized score calculated from a single frame of array I / Q snapshot data in the digital antenna array system. The sensing time can be the moment when the UAV collects relevant data at the sensing waypoint.

[0073] In this embodiment of the invention, when the UAV passes through a sensing waypoint, corresponding feature information and array I / Q snapshot data can be collected at the sensing waypoint.

[0074] The target location can be expressed by the following formula:

[0075] ;

[0076] in, The target location for the drone. The three coordinates are the specific locations of the drone target. This is the matrix transpose.

[0077] Array I / Q snapshot data can be expressed by the following formula:

[0078] ;

[0079] in, This represents the number of time sampling points for array I / Q snapshot data. This represents the array I / Q snapshot data received by the i-th antenna element. This represents the total number of array I / Q snapshots. Due to reflection, diffraction, scattering, occlusion, and noise in complex propagation environments, the array's I / Q snapshot data typically contains a mixed response of direct paths, non-direct paths, and noise components.

[0080] Specifically, the UAV flies according to the sensing waypoints generated during the positioning process, passively receives GNSS interference signals at each sensing moment, and outputs array I / Q snapshot data. In practice, the initial waypoint of the UAV can be given by the mission area or a prior search area, and subsequent waypoints can be updated based on the current positioning results, the validity of the received signal, and the spatial geometric relationship, so that the UAV can gradually obtain multi-view observations that are conducive to positioning.

[0081] For example, after acquiring the feature information and array I / Q snapshot data corresponding to each sensing moment, the feature information and array I / Q snapshot data corresponding to each moment are first processed separately to obtain the corresponding reliable azimuth constraints. Then, multiple reliable azimuth constraints are fused to calculate the fusion weight corresponding to each sensing moment. Based on the fusion weight, the initial position of the interference source when the UAV flies at the sensing waypoint corresponding to the sensing moment can be estimated, which can improve the accuracy of the initial position.

[0082] In one possible approach, the step of determining multi-domain signal representations and physical model angle priors based on array I / Q snapshot data, and calculating reliable azimuth constraints corresponding to each sensing moment based on the target dual-stream neural network, according to the multi-domain signal representations, physical model angle priors, and feature information, includes: Extract the covariance matrix from the array I / Q snapshot data, construct a multi-domain signal representation based on the covariance matrix, and construct a physical model angle prior based on the covariance matrix through electromagnetic propagation mechanism and array receiving model; The target dual-stream neural network is used to process the multi-domain signal representation and the physical model angle prior to obtain the arrival angle probability distribution, confidence level and non-line-of-sight probability. The arrival angle estimate is calculated based on the physical model angle prior and the arrival angle probability distribution. The target position, arrival angle estimate, arrival angle probability distribution, confidence level and non-line-of-sight probability are fused into a reliability-aware angle observation. The reliability weights are calculated based on the reliability perception angle sensing, target position, attitude information, confidence level, non-line-of-sight probability, and received signal effectiveness index. The reliable orientation constraints of the UAV under test are then constructed based on the target position, reliability weights, and angle of arrival estimates.

[0083] It should be understood that, for the feature information and array I / Q snapshot data corresponding to each sensing moment, the covariance matrix is ​​first calculated based on the array I / Q snapshot data. The calculated covariance matrix is ​​a complex-valued matrix, from which corresponding information can be extracted to obtain a multi-domain signal representation. This covariance matrix reflects the correlation between different antenna array elements caused by the incident wavefront. Simultaneously, based on the covariance matrix, a physical model angle prior can be constructed through the electromagnetic propagation mechanism and the array receiving model. The electromagnetic propagation mechanism can be the physical laws governing the propagation of the radio frequency electromagnetic waves radiated by the UAV from the transmitting antenna through the space environment to the receiving antenna array. The array receiving model can be a set of mathematical equations abstracted from the electromagnetic propagation mechanism. The physical model angle prior can be the pre-determined UAV signal incident angle constraint information, combining the electromagnetic wave propagation laws and the array's geometric and physical structure.

[0084] In one possible manner, the extraction of the covariance matrix from the array I / Q snapshot data and the construction of a multi-domain signal representation based on the covariance matrix include:

[0085] Calculate the covariance matrix of the array I / Q snapshot data;

[0086] The covariance matrix is ​​decomposed into real part information, imaginary part information, amplitude information and phase information, and the real part information, imaginary part information, amplitude information and phase information are determined as multi-domain signal representation.

[0087] It should be understood that the covariance matrix is ​​a complex-valued matrix. Therefore, the real part, imaginary part, amplitude, and phase information of the covariance matrix can be extracted and used as a multi-domain signal representation.

[0088] The covariance matrix is ​​expressed by the following formula:

[0089] ;

[0090] in, Let covariance matrix be the variance matrix. For array I / Q snapshot data, This is the conjugate transpose of the array I / Q snapshot data. This represents the number of time sampling points for array I / Q snapshot data;

[0091] Multi-domain signal representation can be expressed by the following calculation formula:

[0092] ;

[0093] in, For multi-domain signal representation, For real information, For imaginary information, For amplitude information, For phase information, It indicates that there are four The three-dimensional real tensor space is formed by concatenating real-valued feature matrices along the channel dimension. This multi-domain signal representation simultaneously preserves the amplitude coupling relationship and phase structure between array elements of the received signal, providing input for subsequent neural networks to extract spatial features under complex propagation environments.

[0094] The prior angles of the physical model are expressed by the following calculation formula:

[0095] ;

[0096] in, As a priori angular feature of the physical model, it can be used to describe the degree of consistency between the candidate angle of arrival direction and the array's physical propagation model. Candidate angle of arrival grid The conjugate transpose of the array guide vector constructed above. The noise subspace is obtained by eigenvalue decomposition of the covariance matrix. This is the conjugate transpose of the noise subspace obtained by eigenvalue decomposition of the covariance matrix. Candidate angle of arrival grid The array guide vector is constructed above. The candidate angle of arrival grid can be a range of achievable angles of arrival; in this embodiment of the invention, the range of candidate angles of arrival is between 0 degrees and 360 degrees.

[0097] The target two-stream neural network can be a classic two-branch parallel convolutional neural network architecture for video temporal data. In embodiments of the present invention, such as Figure 3 As shown, in Figure 3 In this model, the target two-stream neural network is a physics-guided learning two-stream neural network, where B represents the batch size of samples, i.e., the number of samples input into the neural network in each batch. In the specific processing, the target two-stream neural network processes the multi-domain signal representation and the prior angle of the physical model to obtain the angle of arrival probability distribution, confidence level, and non-line-of-sight probability. The angle of arrival probability distribution can be considered as the probability density function corresponding to all possible angles at which the UAV's radio frequency signal is incident on the antenna array. The confidence level is a quantitative indicator calculated based on the angle of arrival probability distribution, used to measure the reliability of the currently calculated signal incident angle. The non-line-of-sight probability is the probability that the electromagnetic wave propagates from the UAV to the antenna array without an unobstructed direct path, relying solely on reflection from walls / ground.

[0098] Then, the estimated angle of arrival can be calculated based on the calculated physical angle prior and the probability distribution of the angle of arrival. Finally, the target position, the estimated angle of arrival, the probability distribution of the angle of arrival, the confidence level, and the non-line-of-sight probability are fused into a reliability-aware angle observation.

[0099] In a possible manner, the target dual-stream neural network includes a signal representation stream network and a physical prior stream network. The process of processing multi-domain signal representations and physical model angle priors through the target dual-stream neural network to obtain the angle of arrival probability distribution, confidence level, and non-line-of-sight probability includes:

[0100] Signal feature information in multi-domain signal representation is extracted by signal representation flow network, and angular structure features in angular prior of physical model are extracted by physical prior flow network.

[0101] The signal feature information and angular structure features are processed to obtain the arrival angle probability distribution, confidence level, and non-line-of-sight probability.

[0102] It should be understood that the target two-stream neural network includes two types of network processing circuits: a signal representation stream network and a physical prior stream network. The signal representation stream network extracts signal feature information from the multi-domain signal representation, including inter-element correlation, amplitude-phase structure, and local spatial coupling features. The physical prior stream network extracts angular structure features from the physical model's angular prior, including peak values, sidelobes, and multi-peak ambiguity in the candidate angle direction. After obtaining the signal feature information and angular structure features, processing these features in the target two-stream neural network yields the arrival angle probability distribution, confidence level, and non-line-of-sight probability.

[0103] Specifically, the probability distribution of the angle of arrival can be expressed by the following formula:

[0104] ;

[0105] in, For the probability distribution of the angle of arrival, This represents the total number of candidate arrival angle directions. For the first One candidate arrival angle, , Candidate arrival angle direction, For the target two-stream neural network in the candidate angle of arrival direction The predicted angle value output above, For the target two-stream neural network in the candidate direction The predicted angle value is output above. The confidence level is... The non-line-of-sight probability is .in, Used to characterize the reliability of the current angle observation. It is used to characterize the probability that the current snapshot is affected by occlusion, reflection, or non-line-of-sight propagation.

[0106] In one possible manner, calculating the estimated angle of arrival based on the physical model angle prior and the angle of arrival probability distribution includes:

[0107] The physical model angle prior is normalized to obtain the normalized physical model angle prior;

[0108] The posterior angle distribution is calculated based on the prior angle and the probability distribution of the angle of arrival from the normalized physical model.

[0109] Extract the estimated angle of arrival from the posterior distribution of the angle.

[0110] It should be understood that normalizing the prior perspective of the physical model into a probabilistic form can be expressed by the following formula:

[0111] ;

[0112] in, The sharpening coefficient is the a priori distribution. For the first k The physical model spatial spectrum after normalization or centering of the candidate arrival angles For the first i The physical model spatial spectrum after normalization or centering of the candidate arrival angles The angle prior is the normalized physical model.

[0113] Then, by dynamically gating and fusing the angle distribution of the neural network with the prior angle of the physical model, the posterior angle distribution is obtained. The posterior angle distribution can be expressed by the following formula:

[0114] ;

[0115] in, For the posterior distribution of angle, For adaptive fusion weights, This refers to the normalized or centered physical model spatial spectrum.

[0116] Based on the fused posterior angle distribution Extract the angle of arrival estimate at the current sensing moment. This can be expressed by the following formula:

[0117] ;

[0118] in, This represents the local angular neighborhood near the peak of the fused angular distribution.

[0119] This forms the reliability perception angle observation at the current perception moment, which can be expressed by the following calculation formula:

[0120] ;

[0121] in, This is an indicator of the validity of the received signal obtained from the original received signal or the detection module.

[0122] After obtaining the reliability-perceived angle observations, reliability weights can be calculated based on the reliability-perceived angle sensing, target position, attitude information, confidence level, non-line-of-sight probability, and received signal effectiveness index. Reliability azimuth constraints for the UAV can then be constructed based on the target position, reliability weights, and estimated angle of arrival. The reliability weights can be used to assign different weights to qualified angle observations, with high weights for good signals and low weights for bad signals. The reliability azimuth constraints can be established as hard boundary conditions for spatial angles based on electromagnetic propagation mechanisms and array physical model angle priors.

[0123] In one possible manner, the calculation of reliability weights based on reliability-aware angle sensing, target position, attitude information, confidence level, non-line-of-sight probability, and received signal effectiveness index includes:

[0124] The direction vector of the UAV in the angle observation is calculated based on the reliability perception angle sensing, target position and attitude information, and the normal vector of the UAV in the angle observation is calculated based on the direction vector.

[0125] Calculate the orthogonal geometric residual based on the normal vector and the target position;

[0126] Reliability weights are calculated based on orthogonal geometric residuals, confidence levels, non-line-of-sight probability, and received signal effectiveness indices.

[0127] It should be understood that, in the specific calculation process, the first step is to determine the horizontal position of the drone within the target location: Then, the direction vector is calculated, which is expressed by the following formula:

[0128] ;

[0129] in, This indicates the spatial orientation corresponding to the current angle of observation. Then, the normal vector is calculated based on the direction vector. This normal vector can be expressed by the following formula.

[0130] ;

[0131] in, This represents the normal vector in that orientation.

[0132] Then, the orthogonal geometric residual is calculated based on the normal vector and the target position. This can be expressed by the following calculation formula:

[0133] ;

[0134] The reliability weight can be expressed by the following formula:

[0135] ;

[0136] in, It is a bounded reliability function. As an indicator of the effectiveness of received signals, Forgetting factor, For the first The geometric residuals corresponding to each observation time increase with increasing confidence and received signal effectiveness, and decrease with increasing non-line-of-sight probability. This transforms the current reliability-awareness perspective observation into a reliability orientation constraint. .

[0137] For example, after processing the array I / Q snapshot data and feature information corresponding to all sensing moments, multiple reliable azimuth constraints can be obtained. The initial position of the interference source corresponding to each sensing moment can be calculated based on the multiple reliable azimuth constraints.

[0138] In one possible manner, calculating the initial position of the interference source corresponding to each sensing moment based on multiple reliable orientation constraints includes: A weighted geometric residual model is constructed based on multiple reliable orientation constraints. The weighted geometric residual model is then solved iteratively using the iterative reweighted least squares method to obtain the fusion weights of the UAV under test at each sensing time. The initial position of the interference source at each sensing time is then calculated based on the fusion weights.

[0139] It should be understood that multiple reliable orientation constraints are combined into a constraint set, and a weighted geometric residual model is constructed based on the constraint set. This weighted geometric residual model can be expressed by the following formula:

[0140] ;

[0141] in, For a weighted geometric residual model, This represents the total number of candidate arrival angle directions. The reliability weight in the reliable orientation constraint. Let be the two-dimensional horizontal position vector of the GNSS interference source to be solved. The target position in the reliable orientation constraint, This is the estimated angle of arrival in the reliable azimuth constraint. This is a vector transpose operation. An index for reliable orientation constraints within a time window. It is a two-dimensional real number space;

[0142] The weighted geometric residual model is solved using the weighted normal equation, which is expressed by the following formula:

[0143] ;

[0144] ;

[0145] ;

[0146] in, For spatial geometric constraint strength, For weighted observation vectors, For regularization terms, It is the normal vector. This is the conjugate transpose of the normal vector;

[0147] To mitigate the impact of anomalous orientation constraints caused by multipath, occlusion, and non-line-of-sight propagation on the localization results, a robust solution is achieved using an iterative reweighted least squares method. First, based on the initial estimate... Calculate the residual of the t-th orientation constraint:

[0148] ;

[0149] The final fusion weights are expressed by the following formula:

[0150] ;

[0151] ;

[0152] in, For the first The fusion weights for each observation time. As a reliability weight, For robust weights, For the first The geometric residuals corresponding to each observation time. This is the residual tolerance parameter; then the initial position of the interference source can be calculated based on the fusion weight corresponding to each sensing time, where the calculation can be solved by the weighted least squares method / weighted equation.

[0153] For example, after obtaining the initial position of the interference source corresponding to each sensing moment, the initial position can be further processed to obtain the final interference source position estimate. First, since the single fusion result may still be affected by instantaneous propagation anomalies, the original position estimates obtained at each sensing moment... An adaptive exponential moving average filter is applied to obtain a smoothed position estimate, which is expressed by the following formula:

[0154] ;

[0155] in, To smooth the position estimation, For the first A smoothed system at each observation time, and , This is the preset minimum smoothing coefficient. This is the preset maximum smoothing coefficient, which can be adjusted based on the current observation reliability and the consistency between the original position estimate and the historical smoothed position estimate; For the first Smooth position estimation at each sensing moment For the first The initial location of the interference source at each sensing moment.

[0156] Subsequently, the target location of the interference source can be determined based on the smoothed position estimate corresponding to each sensing moment, the received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals.

[0157] In one possible manner, determining the target location of the interference source based on the smoothed position estimate corresponding to each sensing time, the received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals includes:

[0158] Multiple sensing moments are sorted in ascending order, and the smoothed position estimate, received signal effectiveness index, spatial geometric constraint, and weighted geometric residual corresponding to each sensing moment are compared with preset conditions. When the smoothed position estimate, received signal effectiveness index, spatial geometric constraint, and weighted geometric residual satisfy the preset conditions, the smoothed position estimate, received signal effectiveness index, spatial geometric constraint, and weighted geometric residual corresponding to each sensing moment in the target sensing moment are determined. The preset conditions are: the spatial geometric constraint is greater than or equal to the spatial geometric constraint strength threshold; the absolute value of the difference between the smoothed position estimate and the smoothed position estimate of the previous sensing moment is less than or equal to the step change threshold; the difference between the smoothed position estimate corresponding to the first sensing moment and the smoothed position estimate corresponding to the last sensing moment within the preset time window is less than or equal to the window drift threshold; the received signal effectiveness index is greater than or equal to the received signal effectiveness threshold; and the difference between the weighted geometric residual and the smoothed position estimate is less than or equal to the preset consistency threshold. The target sensing moment is the current sensing moment and all previous sensing moments.

[0159] For each sensing moment in the target sensing time, a comprehensive quality score is calculated based on the smooth position estimation, received signal effectiveness index, spatial geometric constraints, and weighted geometric residual corresponding to each sensing moment to obtain the initial position of the interference source. After the target sensing time is completed, multiple comprehensive quality scores are obtained.

[0160] Among multiple comprehensive quality scores, the initial position of the interference source corresponding to the highest comprehensive quality score is determined as the target position.

[0161] It should be understood that the initial position, smoothed position estimate, spatial geometric constraints, and received signal effectiveness index at each sensing moment are stored in the historical positioning result set. :

[0162] ;

[0163] Then, a decision can be made based on geometric observability and temporal consistency to determine whether to terminate the localization process. Both geometric observability and temporal consistency are preset conditions. The geometric observability condition of the preset conditions can be expressed by the following formula:

[0164] ;

[0165] in, The threshold for spatial geometric constraint strength. The minimum number of effective sensing moments; `det` is the determinant operator. This represents the strength of spatial geometric constraints. The temporal consistency within the preset conditions can be determined by the step change in the smoothed position estimate within the most recent time window, the window drift, the validity of the received signal, and the consistency between the original estimate and the smoothed estimate. Specifically, this can be expressed by the following formula:

[0166] ;

[0167] ;

[0168] , ;

[0169] in, The threshold for step change, This is the window drift threshold. The threshold for the validity of the received signal. To preset a consistency threshold, For the first The sliding time window corresponding to each observation time. For the first The median is an indicator of the validity of the received signal at each observation time. and The first The and the first Smoothed position estimation for each observation time, For a smoothed position estimate of the first observation time within the sliding time window, This is a 2-norm operator.

[0170] The positioning process terminates when preset conditions are met or the maximum sensing budget is reached. After termination, instead of directly using the position estimate from the last sensing moment, a candidate position with good temporal consistency, small weighted geometric residual, and strong spatial geometric constraints is selected from the target sensing moments as the final GNSS interference source position estimate. Specifically, a comprehensive quality score is calculated for each sensing moment in the target sensing time, resulting in multiple comprehensive quality scores. The initial position of the interference source corresponding to the highest comprehensive quality score is then determined as the target position.

[0171] In some possible ways, the positioning method further includes:

[0172] When the smoothed position estimation, received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals do not meet the preset conditions, the estimated angle of arrival and initial position corresponding to the sensing moment are obtained, and the position of the sensing waypoint corresponding to the next sensing moment is updated based on the estimated angle of arrival and initial position.

[0173] It should be understood that if preset conditions are not met, the position of the sensing waypoint at the next sensing moment can be updated based on the estimated angle of arrival and the initial position. Specifically, such as... Figure 4 As shown, where, , and The radial direction is used. The direction of movement of the UAV can be obtained by combining the radial and tangential components. The radial direction is beneficial for enhancing the received signal, while the tangential direction is beneficial for increasing the diversity of azimuth intersection angles. Let the current radial and tangential directions be:

[0174] ;

[0175] ;

[0176] in, For the first A unit radial vector along the current angle of arrival at each sensing moment. Let be the unit tangential vector orthogonal to the radial vector. The direction of the UAV's next horizontal movement can be represented as:

[0177] ;

[0178] ;

[0179] in, Weights used to control radial and tangential motion. Used to introduce a lateral bias that varies over time. This is the synthesized unnormalized horizontal motion direction vector. This is the normalized unit vector for the next horizontal movement direction. This is the lateral offset amplitude coefficient. This is the phase variable used to control the periodic variation of the lateral offset. The UAV's horizontal position is updated as follows:

[0180] ;

[0181] in, Waypoint step size, Let be the two-dimensional horizontal coordinates of the UAV at the t-th waypoint. Through the above waypoint update method, the UAV can continuously acquire new azimuth observations during the positioning process and gradually improve the spatial geometry conditions for multi-time azimuth fusion.

[0182] In practical applications, existing positioning methods exhibit significant accuracy degradation under complex multipath propagation scenarios, occlusion, and non-line-of-sight conditions, with some positioning results deviating considerably from the actual interference source location. In contrast, the method of this invention achieves positioning results closer to the true location in all scenarios and effectively reduces abnormal positioning results. Therefore, this invention, through physically guided angle observation extraction and multi-moment spatial geometric fusion, can improve the accuracy and stability of GNSS interference source positioning under the influence of reflection, diffraction, scattering, occlusion, and noise.

[0183] The above technical solution introduces the concept of physical-guided learning, combining electromagnetic propagation mechanisms, array spatial response, and multi-domain array I / Q snapshot representation. It also uses a dual-stream neural network to jointly process signal features and physical model angle priors. This allows the network to learn signal distortion characteristics under complex environments while being constrained by physical priors, thereby improving the accuracy and robustness of angle-of-arrival estimation under the influence of reflection, diffraction, scattering, obstruction, and noise. Utilizing the spatial geometric constraints formed by the continuous movement of the UAV, robust temporal fusion of angle-of-arrival observations from multiple sensing moments is achieved, suppressing inconsistent observations caused by multipath and non-line-of-sight propagation and improving the stability of interference source location estimation. Only one UAV equipped with a small antenna array is needed to locate GNSS interference sources, requiring minimal observation data, incurring low operating costs, and offering flexible deployment. This facilitates engineering implementation in GNSS interference source investigation in complex urban environments, airport perimeters, ports, and key areas. This invention achieves robust localization of GNSS interference sources in complex propagation environments by acquiring array I / Q snapshot data, multi-domain signal characterization, physical model angle prior, physically guided target dual-stream neural network, reliable azimuth constraints, multi-time azimuth fusion, and temporal position smoothing and historical optimal output.

[0184] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A robust localization method for UAV-borne GNSS interference sources based on physics-guided learning, characterized in that, include: The array I / Q snapshot data of the UAV at each sensing moment and the feature information of the UAV at each sensing moment are acquired. The feature information includes the target position, attitude information and received signal validity index of the UAV. The UAV is equipped with an antenna array. For each sensing moment, the array I / Q snapshot data and feature information are used to determine the multi-domain signal representation and physical model angle prior based on the array I / Q snapshot data. Based on the target dual-stream neural network, the reliable orientation constraint corresponding to each sensing moment is calculated according to the multi-domain signal representation, physical model angle prior and feature information. After processing all array I / Q snapshot data and feature information corresponding to each sensing moment, multiple reliable azimuth constraints are obtained, and the initial position of the interference source corresponding to each sensing moment is calculated based on the multiple reliable azimuth constraints. For the initial position of the interference source corresponding to each sensing time, the following processing is performed: the initial position of the interference source is processed by adaptive exponential moving average filtering to obtain a smooth position estimate, and the spatial geometric constraints and weighted geometric residuals corresponding to each sensing time are determined; The target location of the interference source is determined based on the smoothed position estimate corresponding to each sensing moment, the received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals.

2. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 1, characterized in that, The process of determining multi-domain signal representations and physical model angle priors based on array I / Q snapshot data, and calculating reliable azimuth constraints corresponding to each sensing moment based on the target dual-stream neural network, using multi-domain signal representations, physical model angle priors, and feature information, includes: Extract the covariance matrix from the array I / Q snapshot data, construct a multi-domain signal representation based on the covariance matrix, and construct a physical model angle prior based on the covariance matrix through electromagnetic propagation mechanism and array receiving model; The target dual-stream neural network is used to process the multi-domain signal representation and the physical model angle prior to obtain the arrival angle probability distribution, confidence level and non-line-of-sight probability. The arrival angle estimate is calculated based on the physical model angle prior and the arrival angle probability distribution. The target position, arrival angle estimate, arrival angle probability distribution, confidence level and non-line-of-sight probability are fused into a reliability-aware angle observation. The reliability weights are calculated based on the reliability perception angle sensing, target position, attitude information, confidence level, non-line-of-sight probability, and received signal effectiveness index. The reliable orientation constraints of the UAV under test are then constructed based on the target position, reliability weights, and angle of arrival estimates.

3. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 1, characterized in that, The calculation of the initial position of the interference source corresponding to each sensing moment based on multiple reliable orientation constraints includes: A weighted geometric residual model is constructed based on multiple reliable orientation constraints. The weighted geometric residual model is then solved iteratively using the iterative reweighted least squares method to obtain the fusion weights of the UAV under test at each sensing time. The initial position of the interference source at each sensing time is then calculated based on the fusion weights.

4. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 1, characterized in that, The process of determining the target location of the interference source based on the smoothed position estimate, received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals corresponding to each sensing time includes: Multiple sensing moments are sorted in ascending order, and the smoothed position estimate, received signal effectiveness index, spatial geometric constraint, and weighted geometric residual corresponding to each sensing moment are compared with preset conditions. When the smoothed position estimate, received signal effectiveness index, spatial geometric constraint, and weighted geometric residual satisfy the preset conditions, the smoothed position estimate, received signal effectiveness index, spatial geometric constraint, and weighted geometric residual corresponding to each sensing moment in the target sensing moment are determined. The preset conditions are: the spatial geometric constraint is greater than or equal to the spatial geometric constraint strength threshold; the absolute value of the difference between the smoothed position estimate and the smoothed position estimate of the previous sensing moment is less than or equal to the step change threshold; the difference between the smoothed position estimate corresponding to the first sensing moment and the smoothed position estimate corresponding to the last sensing moment within the preset time window is less than or equal to the window drift threshold; the received signal effectiveness index is greater than or equal to the received signal effectiveness threshold; and the difference between the weighted geometric residual and the smoothed position estimate is less than or equal to the preset consistency threshold. The target sensing moment is the current sensing moment and all previous sensing moments. For each sensing moment in the target sensing time, a comprehensive quality score is calculated based on the smooth position estimation, received signal effectiveness index, spatial geometric constraints, and weighted geometric residual corresponding to each sensing moment to obtain the initial position of the interference source. After the target sensing time is completed, multiple comprehensive quality scores are obtained. Among multiple comprehensive quality scores, the initial position of the interference source corresponding to the highest comprehensive quality score is determined as the target position.

5. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 4, characterized in that, The positioning method further includes: When the smoothed position estimation, received signal effectiveness index, spatial geometric constraints, and weighted geometric residuals do not meet the preset conditions, the estimated angle of arrival and initial position corresponding to the sensing moment are obtained, and the position of the sensing waypoint corresponding to the next sensing moment is updated based on the estimated angle of arrival and initial position.

6. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 2, characterized in that, The extraction of the covariance matrix from the array I / Q snapshot data, and the construction of a multi-domain signal representation based on the covariance matrix, includes: Calculate the covariance matrix of the array I / Q snapshot data; The covariance matrix is ​​decomposed into real part information, imaginary part information, amplitude information and phase information, and the real part information, imaginary part information, amplitude information and phase information are determined as multi-domain signal representation.

7. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 6, characterized in that, The covariance matrix is ​​expressed by the following formula: ； in, Let covariance matrix be the variance matrix. For array I / Q snapshot data, This is the conjugate transpose of the array I / Q snapshot data. This represents the number of time sampling points for array I / Q snapshot data; The prior angle of the physical model is expressed by the following calculation formula: ； in, As a priori from the perspective of the physical model, This is the conjugate transpose of the array guide vector. The noise subspace is obtained by eigenvalue decomposition of the covariance matrix. This is the conjugate transpose of the noise subspace obtained by eigenvalue decomposition of the covariance matrix. The array guide vector for the candidate direction.

8. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 2, characterized in that, The target dual-stream neural network includes a signal representation stream network and a physical prior stream network. The process of using the target dual-stream neural network to process multi-domain signal representations and physical model angle priors yields the arrival angle probability distribution, confidence level, and non-line-of-sight probability, including: The signal feature information in the multi-domain signal representation is extracted by the signal representation flow network, and the angle structure features in the angle prior of the physical model are extracted by the physical prior flow network. The signal feature information and angular structure features are processed to obtain the arrival angle probability distribution, confidence level, and non-line-of-sight probability.

9. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 2, characterized in that, The calculation of the estimated angle of arrival based on the prior angle of the physical model and the probability distribution of the angle of arrival includes: The physical model angle prior is normalized to obtain the normalized physical model angle prior; The posterior angle distribution is calculated based on the prior angle and the probability distribution of the angle of arrival from the normalized physical model. Extract the estimated angle of arrival from the posterior distribution of the angle.

10. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 9, characterized in that, The posterior distribution of the angle is expressed by the following formula: ； ； ； in, For the posterior distribution of angle, For the probability distribution of the angle of arrival, For adaptive fusion weights, For the normalized physical model perspective prior, For the first k The physical model spatial spectrum after normalization or centering of the candidate arrival angles For the first i The physical model spatial spectrum after normalization or centering of the candidate arrival angles The sharpening coefficient is the a priori distribution. This represents the total number of candidate arrival angle directions. For the first One candidate arrival angle, , Candidate arrival angle direction, For the target two-stream neural network in the candidate angle of arrival direction The predicted angle value output above; For the target two-stream neural network in the candidate direction The predicted angle value output above.

11. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 2, characterized in that, The calculation of reliability weights based on reliability-perceived angle sensing, target position, attitude information, confidence level, non-line-of-sight probability, and received signal effectiveness index includes: The direction vector of the UAV in the angle observation is calculated based on the reliability perception angle sensing, target position and attitude information, and the normal vector of the UAV in the angle observation is calculated based on the direction vector. Calculate the orthogonal geometric residual based on the normal vector and the target position; Reliability weights are calculated based on orthogonal geometric residuals, confidence levels, non-line-of-sight probability, and received signal effectiveness indices.

12. The robust localization method for UAV-borne GNSS interference sources based on physics-guided learning according to claim 3, characterized in that, The weighted geometric residual model is expressed by the following formula: ； in, For a weighted geometric residual model, This represents the total number of candidate arrival angle directions. The reliability weight in the reliable orientation constraint. Let be the two-dimensional horizontal position vector of the GNSS interference source to be solved. The target position in the reliable orientation constraint. This is the estimated angle of arrival in the reliable azimuth constraint. This is a vector transpose operation. This is an index for reliable orientation constraints within a time window. It is a two-dimensional real number space; The weighted geometric residual model is solved using a weighted method equation, which is expressed by the following formula: ； ； ； in, For spatial geometric constraint strength, For weighted observation vectors, For regularization terms, It is the normal vector. This is the conjugate transpose of the normal vector; The fusion weight is expressed by the following formula: ； ； in, For the first The fusion weights for each observation time. For reliability weights, For robust weights, For the first The geometric residuals corresponding to each observation time. This is the residual tolerance parameter; The smoothed position estimate is expressed by the following formula: ； in, To smooth the position estimation, For the first The smoothing coefficient at each observation time. For the first Smooth position estimation at each sensing moment For the first The initial location of the interference source at each observation time.

Images