A GNSS anti-jamming closed-loop phase self-correction method based on deep learning

Abstract

The application discloses a GNSS anti-interference closed-loop phase self-correction method based on deep learning, relates to the technical field of satellite navigation and artificial intelligence, and acquires an anti-interference array output signal and weight information through an array antenna and space-time adaptive processing; extracts a carrier-to-noise ratio, weight phase feature and historical positioning quality, and constructs a deep learning feature vector; utilizes a multilayer perception machine model to infer an array element level phase correction reference quantity, and realizes accurate phase compensation of the array output signal by combining phase error layered modeling, fast-changing disturbance prediction and weight sensitivity estimation; introduces an anti-interference perception observation covariance dynamic scaling mechanism in RTK positioning calculation; and triggers online learning and model updating through positioning quality evaluation results, so that the system has adaptive evolution capability. The application can significantly improve GNSS carrier phase quality and positioning precision under strong interference conditions, realizes deep fusion of anti-interference processing and high-precision positioning, and has important engineering application value.

Description

Technical Field

[0001] This invention relates to the fields of satellite navigation and artificial intelligence technology, and in particular to a GNSS anti-interference closed-loop phase self-correction method based on deep learning. Background Technology

[0002] With the increasing density of various ground-based communication base stations, surveillance radars, and security / counter-electromagnetic equipment, the electromagnetic interference on satellite navigation signals is becoming increasingly stronger in both intensity and frequency. To improve anti-interference performance, GNSS receivers typically integrate array antennas and anti-interference technologies such as space-time adaptive processing (STAP), beamforming, and spatial interference suppression to significantly improve signal reception quality and anti-interference capabilities.

[0003] However, existing anti-interference methods generally suffer from a difficult-to-resolve technical contradiction: changes in anti-interference weights directly introduce additional array phase distortion, causing the GNSS carrier phase to no longer satisfy the phase continuity assumption of traditional positioning models, resulting in a significant decrease in differential positioning accuracy. Especially in scenarios with strong interference, frequent weight updates, and low signal-to-noise ratios, receiver carrier phase observations may exhibit abrupt changes, drift, and additional noise, making it difficult to correctly fix the integer ambiguity of RTK high-precision positioning methods, resulting in the technical contradiction of "anti-interference but decreased accuracy." Traditional filtering techniques cannot quantify the impact of anti-interference weights on the phase in real time, lack element-level phase compensation mechanisms, and struggle to distinguish between instantaneous phase jitter caused by rapidly changing interference and slowly varying offset caused by hardware drift. Summary of the Invention

[0004] To overcome the shortcomings of the existing technology, this invention provides a GNSS anti-interference closed-loop phase self-correction method based on deep learning. It constructs a closed-loop system from anti-interference output to phase repair, then to high-precision positioning, and finally to online learning. It aims to solve the problems of severe carrier phase distortion, difficulty in fixing RTK ambiguity, and significant decrease in positioning accuracy after array anti-interference processing.

[0005] To achieve the above objectives, the present invention adopts the following technical solution, including:

[0006] A deep learning-based GNSS anti-interference closed-loop phase self-calibration method includes the following steps:

[0007] S1. Obtain the complex baseband signal of each element of the array antenna, and use the space-time adaptive processing anti-interference algorithm to obtain the anti-interference array output signal vector. With anti-interference weight vector ; where subscript Indicates an epoch, or a point in time;

[0008] S2. Constructing signal quality feature vectors And extract the anti-interference weight phase feature vector. and historical positioning quality feature vector The feature vectors are concatenated to obtain the input feature vectors of the deep learning model. ;

[0009] S3, input feature vector Inputting the deep learning model yields the phase correction reference vector. ;

[0010] S4. Decompose the measured phase residual to obtain the phase slowly varying offset. and phase rapid drift The phase rapid change drift prediction is obtained by using the phase rapid change drift of the most recent w epochs. ;

[0011] S5, Based on phase correction reference vector Phase slow-varying bias and phase rapid change drift prediction The phase compensation vector is obtained. Used to counter interference array output signal vector Apply phase correction to obtain the phase-corrected array output signal vector. ;

[0012] S6, based on the phase-corrected array output signal vector Perform RTK positioning calculations to obtain the positioning results. With positioning quality indicators ;

[0013] S7. Confidence level is determined based on ambiguity fixation rate and phase noise index; quality is evaluated based on positioning quality index; and data pairs meeting the set confidence criteria are selected as sample pairs. The deep learning model is trained online when the positioning quality meets the set conditions.

[0014] Preferably, step S1 is as follows:

[0015] S11. Receive the original radio frequency time domain signal from each element of the array antenna. The original radio frequency time domain signal contains satellite navigation signal, interference signal and noise. Perform down-conversion, bandpass filtering and gain control on the original radio frequency time domain signal of each element to shift the effective components in the working frequency band to intermediate frequency or baseband to obtain the complex intermediate frequency / complex baseband signal of each element.

[0016] S12. Perform analog-to-digital conversion on the complex intermediate frequency / complex baseband signals of each array element to obtain the discrete complex baseband sampling sequence of each array element; in each epoch... The complex baseband sampled values ​​of each array element are arranged into an array data vector according to the array element index. :

[0017] ;

[0018] in, Indicates the first Individual elements in the historical period The complex baseband sampled values; the total number of array elements is ;

[0019] S13. Within a preset time or snapshot window, based on the array data vector Estimate the interference plus noise covariance matrix covariance matrix The diagonal elements represent the received power of each array element, and the off-diagonal elements represent the spatial correlation between array elements.

[0020] S14. Based on the estimated interference plus noise covariance matrix and the steering vector of the desired satellite signal An optimization problem for space-time adaptive processing is constructed, which minimizes interference and noise power while satisfying the desired satellite signal gain constraint, and the anti-interference weight vector is obtained by solving the problem. :

[0021] ;

[0022] ;

[0023] in, Indicates the first The weight range of each array element, Indicates the first Individual elements in the historical period The weighted phase adjustment amount is the weighted phase characteristic. Indicates the epoch For the first Unit amplitude complex phase rotation applied to each array element channel, Indicates the first Individual elements in the historical period The anti-interference weight, where j represents the imaginary unit;

[0024] S15. Utilizing anti-interference weight vectors For array data vectors Weighted synthesis is performed to obtain the output signal vector of the anti-interference array. :

[0025] ;

[0026] in, Indicates the first Individual elements in the historical period The complex baseband signal after space-time adaptive processing, i.e., the first... Individual elements in the historical period The anti-interference output signal .

[0027] Preferably, step S2 is as follows:

[0028] S21, Output signal vector of the anti-jamming array Conduct signal quality assessments for each satellite. The carrier-to-noise ratio is calculated using the relevant output power and noise estimates, thus constructing a signal quality feature vector. :

[0029] ;

[0030] in, Indicates the first A satellite in the era The carrier-to-noise ratio; Indicates the total number of satellites; This indicates the useful signal power of the relevant output. This represents the equivalent noise power spectral density or noise estimate.

[0031] S22. Extract the phase features of the anti-interference weights from the anti-interference weight vector. Extract the weighted phase feature of each array element, i.e., the weighted phase adjustment amount. This constitutes the anti-interference weighted phase eigenvector. :

[0032] ;

[0033] The total number of array elements is ;

[0034] S23. Obtain the ambiguity fixation rate of the previous epoch from the RTK positioning solution module. and phase noise index This constitutes the historical positioning quality feature vector. :

[0035] ;

[0036] S24. Concatenate the three feature vectors obtained in a predetermined order to form the input feature vector of the deep learning model:

[0037] ;

[0038] in, This represents the concatenation operation of eigenvectors; This represents the concatenated feature vector;

[0039] Normalization or standardization is performed on the concatenated feature vectors to bring all features into a uniform dimension and range, thus forming the final input feature vector. .

[0040] Preferably, step S3 is as follows:

[0041] S31. Input feature vector Deep learning models input into a multilayer perceptron structure Linear mapping and activation operations are performed sequentially according to the network hierarchy:

[0042] ;

[0043] The output of the deep learning model is obtained through a linear mapping of the output layer. ;

[0044] in, , This represents the weight matrix and bias vector of the i-th layer; This represents the hidden layer output; the deep learning model has a total of s layers; This represents the linear rectification function, also known as the activation function. Indicates the current epoch. Model parameters;

[0045] S32, the output of the deep learning model is The phase correction reference quantity of each array element is used to obtain the phase correction reference vector. ;in, The first inference obtained from the model represents the... Phase correction reference quantity for each array element;

[0046] S33. To ensure that the output of the deep learning model does not compromise the stability of subsequent phase compensation, an optional consistency check is performed: First, an amplitude limit is applied to the amplitude, with the following condition: , The first step is to set an amplitude threshold; the second step is to apply stability constraints to changes in adjacent epochs, with the following constraints: , This is the threshold for the amount of change; if the constraint is triggered, the output of the deep learning model is pruned to obtain the final phase correction reference vector. .

[0047] Preferably, step S4 is as follows:

[0048] S41. Extract the measured carrier phase observation value of the current epoch from the RTK observation generation module. The expected geometric phase is calculated based on the satellite geometric model and the coordinate prediction values ​​from the previous epoch. Calculate the measured phase residual ;

[0049] S42, Based on the measured phase residual The time-varying characteristics of the phase residual divide it into a slowly varying bias term and a rapidly varying drift term. A Kalman filter is used to extract the slowly varying bias term to obtain the slowly varying phase bias. :

[0050] ;

[0051] Thus, the phase rapid shift amount is obtained. :

[0052] ;

[0053] in, Indicates the current epoch. Previous historical epoch index; The set of windows representing the Kalman filter; Indicates the historical era Measured phase residual at the location; Represents the expectation operator;

[0054] S43. Based on the anti-interference weight vector of the current epoch. Anti-interference weight vector of the previous epoch Calculate the weight increment , Characterizes the weight adjustment of the spatiotemporal adaptive processing anti-interference algorithm due to changes in the current interference environment;

[0055] S44. Construct an approximate linear relationship between the weight increment and the measured phase residual. ,in, For the current epoch The weight-phase sensitivity vector is used to describe the degree of influence of weight changes on the output phase. A hybrid algorithm combining the Kalman filter algorithm and the expectation-maximization algorithm is used to estimate the weight-phase sensitivity vector online. ; Represents the weight-phase sensitivity vector;

[0056] S45. Construct a sequence from the phase rapid shifts of the most recent w epochs. will sequence Input a time series prediction model to predict the current epoch. Phase fast drift prediction :

[0057] ;

[0058] in, This represents a time series prediction model.

[0059] Preferably, step S5 is as follows:

[0060] S51, Change the phase slowly by the bias amount As a slowly varying compensation quantity ;Predict the phase rapid change drift As a rapid change compensation quantity According to the weight-phase sensitivity vector Compared with the current weight increment Calculate the phase change estimate ;

[0061] S52, Phase correction reference vector Slow-varying compensation amount Rapidly changing compensation amount Phase change estimator Weighted fusion is performed to form the final phase compensation vector:

[0062] ;

[0063] in, For weighting;

[0064] ;

[0065] in, Indicates the first Individual elements in the historical period Phase compensation amount; total number of array elements is ;

[0066] S53, anti-jamming array output signal vector By applying a complex phase rotation to each array element, the phase-corrected array element output signal is obtained. This forms the phase-corrected array output signal vector:

[0067] ;

[0068] in, Represents the output signal vector of the anti-interference array The Middle The anti-interference output signal of each array element.

[0069] Preferably, in step S6, a double-difference observation model is used in the RTK positioning solution process, and a dynamic scaling mechanism for observation covariance is introduced to fix ambiguity, as detailed below:

[0070] S61. Based on the signal quality feature vector and anti-interference weight phase eigenvector and phase compensation amount Construct a weighted adaptive covariance matrix ;

[0071] S62. In the double-difference observation model, the weighted adaptive covariance matrix is ​​used. Phase-corrected array output signal vector Weighted calculations are performed, and combined with fixed ambiguity, to obtain the positioning result. Output quality indicators ;in, Indicates the current epoch. The ambiguity fixation rate; Indicates the current epoch. Phase noise index; Indicates the current epoch. The compensation consistency index is determined by the phase compensation change between adjacent epochs. Decide; This represents the satellite's geometric parameters, determined by the geometric distribution parameters output by the receiver.

[0072] Preferably, step S7 is as follows:

[0073] S71. The high confidence criterion is: the ambiguity fixation rate satisfies... And the phase noise index meets ,in, , The thresholds for ambiguity fixation rate and phase noise index are used; if the high confidence criterion is met, then a sample pair for online learning is formed. ;

[0074] S72. Quality evaluation criteria are: the positioning quality indicators meet the requirements. ,in, The threshold for locating quality indicators is set; if the quality evaluation criteria are met, online learning is triggered and the model parameters are updated; otherwise, online learning is not triggered and the model parameters are not updated.

[0075] S73, using online learning samples to... Define supervised learning loss :

[0076]

[0077] in, This represents the output of a deep learning model; These are the model parameters for the current epoch;

[0078] The stochastic gradient descent optimizer is used to update the model parameters based on the loss gradient.

[0079]

[0080] in, For online learning rate, Represents the loss gradient;

[0081] S74. Monitor the ambiguity fixation rate within a consecutive epoch after the deep learning model is updated. Phase noise index Phase compensation change And positioning residual index, If any of the following occurs: ambiguity fixation rate decreases beyond the set range; phase noise index increases beyond the set range; phase compensation change exceeds the set value; or positioning residual index exceeds the set value, then the update of the model parameters is deemed to have a negative effect. The model rollback mechanism is then executed to restore the model parameters to their values ​​before the update. Use these as the current optimal model parameters and clear the sample pairs from the most recent period, i.e., failed sample pairs; otherwise, retain them. As the current optimal model parameters;

[0082] S75. For successful sample pairs that have not been cleared, add them to the experience pool;

[0083] S76. Use the current optimal model parameters as the model parameters for the next epoch.

[0084] The present invention also provides an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned deep learning-based GNSS anti-interference closed-loop phase self-calibration method.

[0085] The present invention also provides a computer program product comprising a computer program / instruction that, when executed by a processor, implements the aforementioned deep learning-based GNSS anti-interference closed-loop phase self-calibration method.

[0086] The advantages of this invention are:

[0087] (1) Existing array anti-interference methods typically suppress interference sources through algorithms such as STAP, MVDR, and PI. However, these weight updates inevitably introduce array element-level phase distortion, causing carrier phase observations to no longer satisfy the continuity assumption of the RTK model, resulting in a significant decrease in ambiguity fixation rate. Existing technologies mainly rely on analytical modeling, fixed filtering, or simple smoothing to compensate for phase deviations. They are difficult to simultaneously handle the instantaneous rapid error caused by weight changes and the slow bias caused by hardware drift, and they cannot adapt to complex dynamic interference environments. This invention proposes a deep learning-driven array element-level phase reference inference method, which can learn the nonlinear mapping relationship between weight changes and phase distortion using only observables inside the receiver (carrier-to-noise ratio, anti-interference weight phase, and historical positioning quality). At the same time, it combines phase error hierarchical modeling and weight sensitivity terms to achieve high-precision recovery of complex phase disturbances.

[0088] (2) Existing anti-interference positioning technologies are mostly open processing links, and the anti-interference module and RTK positioning module lack interaction, making it impossible to form effective feedback. This invention innovatively constructs a closed-loop structure of "anti-interference - phase compensation - RTK positioning quality feedback - online learning". It enhances the RTK solution capability through the dynamic scaling mechanism of observation covariance of anti-interference perception, and uses indicators such as ambiguity fixation rate, phase noise, compensation consistency and satellite geometry as confidence control signals for model learning. This enables the system to perform self-distillation-style online updates at high quality moments, and ensures the safety and reliability of updates through a rollback mechanism, thereby realizing the adaptive evolution of the anti-interference positioning system.

[0089] (3) In the prior art, deep learning is mainly used for signal detection, interference identification, and multipath mitigation, but not for array element-level phase compensation, and no system has been formed that couples deep learning with RTK ambiguity fixation in a closed loop. In this invention, the phase correction reference quantity output by the deep learning model is integrated with slow-varying offset compensation and fast-varying disturbance prediction to form a high-precision phase recovery mechanism that can run in real time. This effectively eliminates the damage to RTK performance caused by array weight updates, and significantly improves the ambiguity fixation rate and positioning stability under strong interference conditions. Attached Figure Description

[0090] Figure 1 This is a flowchart of the method of the present invention.

[0091] Figure 2 This is an experimental result diagram of the present invention (positioning accuracy under anti-interference conditions). Detailed Implementation

[0092] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0093] Depend on Figure 1 As shown, a deep learning-based GNSS anti-interference closed-loop phase self-calibration method includes the following steps:

[0094] S1. By receiving and processing the array signal through the anti-interference module, the complex baseband signal of each element of the array antenna is obtained, and the output signal vector of the anti-interference array is obtained by using the space-time adaptive processing anti-interference algorithm. With anti-interference weight vector .

[0095] Step S1 is as follows:

[0096] S11. Receive the original radio frequency time domain signals from each element of the array antenna. The original radio frequency time domain signals simultaneously contain satellite navigation signals, interference signals, and noise. Perform down-conversion, bandpass filtering, and gain control on the original radio frequency time domain signals of each element respectively, and shift the effective components in the BeiDou / GNSS working frequency band to the intermediate frequency or baseband to obtain the complex intermediate frequency / complex baseband signals of each element.

[0097] S12. Perform analog-to-digital conversion on the analog complex intermediate frequency / complex baseband signals of each array element to obtain the discrete complex baseband sampling sequence of each array element; at each epoch (time point). The complex baseband sampled values ​​of each array element are organized into an array data vector according to the element index, providing input for subsequent space-time adaptive processing. As shown below:

[0098] ;

[0099] in, Indicates the first Individual elements in the historical period The complex baseband sampled values; in this embodiment, a 4-element matrix array is selected, that is, the total number of array elements is... =4.

[0100] Array data vector Each element in The current complex signal value corresponding to an array element includes complex form signals of satellite navigation signals, interference signals, and noise:

[0101] ;

[0102] in, Indicates the first Individual elements in the historical period The in-phase component reflects the projection of the signal onto the cosine (cos) basis direction; Indicates the first Individual elements in the historical period The orthogonal components reflect the projection of the signal onto the sinusoidal basis direction; It represents the imaginary unit.

[0103] S13. Within a preset time or snapshot window, based on the array data vector Estimate the interference plus noise covariance matrix covariance matrix The diagonal elements represent the received power of each array element, while the off-diagonal elements represent the spatial correlation between array elements, providing statistical information for solving the space-time adaptive anti-interference weights.

[0104] S14. Based on the estimated interference plus noise covariance matrix and the steering vector of the desired satellite signal An optimization problem for space-time adaptive processing is constructed, which minimizes interference and noise power while satisfying the desired satellite signal gain constraint, and the anti-interference weight vector is obtained by solving the problem. :

[0105] ;

[0106] ;

[0107] in, Indicates the first The weight range of each array element, Indicates the first Individual elements in the historical period The weighted phase adjustment amount is the weighted phase characteristic. Indicates the epoch For the first Unit amplitude complex phase rotation applied to each array element channel, Indicates the first Individual elements in the historical period Anti-interference weights.

[0108] S15. Utilizing anti-interference weight vectors For array data vectors Weighted synthesis is performed to obtain the output signal vector of the anti-interference array. :

[0109] ;

[0110] in, Indicates the first Individual elements in the historical period The complex baseband signal after space-time adaptive processing, i.e., the first... Individual elements in the historical period The anti-interference output signal .

[0111] Each array element in the epoch Weighted phase adjustment amount That is, the weighted phase features are stored so that the anti-interference weighted phase feature vector can be extracted in the subsequent step S2. use.

[0112] The obtained anti-interference array output signal vector This serves as the input for subsequent carrier tracking and observation generation. The anti-interference weight vector... This serves as input for subsequent deep learning feature construction and phase error modeling, thereby bridging the gap between anti-interference and closed-loop phase self-correction.

[0113] S2. Constructing signal quality feature vectors And extract the phase component of the anti-interference weight. And the historical positioning quality feature vector formed by the ambiguity fixation rate and phase noise index of the previous epoch. The three feature vectors are concatenated to obtain the input feature vector of the deep learning model.

[0114] Step S2 is as follows:

[0115] S21, Output signal vector of the anti-jamming array Conduct signal quality assessments for each satellite. The carrier-to-noise ratio is calculated using the relevant output power and noise estimates, thus constructing a signal quality feature vector. :

[0116] ;

[0117] in, Indicates the first A satellite in the era The carrier-to-noise ratio is used to reflect the degree to which the signals of each satellite are preserved after interference suppression; Indicates the total number of satellites; This indicates the useful signal power of the relevant output. This represents the equivalent noise power spectral density or noise estimate.

[0118] S22. Extract the anti-interference weight phase features from... Extract the weighted phase feature of each array element, i.e., the weighted phase adjustment amount. This constitutes the anti-interference weighted phase eigenvector. :

[0119] ;

[0120] This is used to reflect the one-time impact of changes in the space-time adaptive anti-interference weights on the array output phase.

[0121] S23. Obtain the ambiguity fixation rate of the previous epoch from the RTK positioning solution module. and phase noise index This constitutes the historical positioning quality feature vector. :

[0122] ;

[0123] in, This reflects the reliability of the fixed ambiguity in the previous epoch. Characterizes the phase stability of carrier observations.

[0124] S24. Perform feature splicing and unified formatting:

[0125] The three feature vectors obtained are concatenated in a predetermined order to form the input feature vector of the deep learning model:

[0126] ;

[0127] in, This represents the concatenation operation of eigenvectors; This represents the concatenated feature vector;

[0128] The concatenated feature vectors are then normalized or standardized to ensure that all features are within a uniform dimension and range, thus forming the final input feature vector. .

[0129] S3. Based on the phase correction reference quantity of the inference array element of the deep learning model, the input feature vector is input into the deep learning model of the multilayer perceptron structure, and the phase correction reference vector is obtained through forward propagation.

[0130] Step S3 is as follows:

[0131] S31. Input feature vector Input to a deep learning model with a multilayer perceptron (MLP) structure Linear mapping and activation operations are performed sequentially according to the network hierarchy:

[0132] ;

[0133] The output of the deep learning model is obtained through a linear mapping of the output layer. ;

[0134] in, , This represents the weight matrix and bias vector of the i-th layer; This represents the hidden layer output; the deep learning model has a total of s layers; This represents the linear rectification function, also known as the activation function. This represents the model parameters for the current epoch.

[0135] S32. The output of the deep learning model is... The phase correction reference quantity (phase compensation reference quantity) of each array element is used to obtain the phase correction reference vector. ;

[0136] in, The first inference obtained from the model represents the... The phase correction reference value for each array element is used to compensate for array phase disturbances introduced by space-time adaptive processing.

[0137] S33. To ensure that the output of the deep learning model does not compromise the stability of subsequent phase compensation, an optional consistency check is performed on the model's output: First, an amplitude limit is applied to the amplitude, with the following condition: , The first step is to set an amplitude threshold; the second step is to apply stability constraints to changes in adjacent epochs, with the following constraints: , The threshold value represents the amount of change. If any of the above restrictions are triggered, the output of the deep learning model is pruned to obtain the final phase correction reference vector after limiting. .

[0138] S4. Perform phase error hierarchical modeling and short-time dynamic prediction. Decompose the measured phase residual into a slow-varying bias term and a fast-varying drift term to obtain the phase slow-varying bias amount and the phase fast-varying drift amount. Then, use a lightweight time-series prediction model to make a short-time prediction of the phase fast-varying drift amount in the nearest window to obtain the phase fast-varying drift prediction amount.

[0139] Step S4 is as follows:

[0140] S41. Extract the measured carrier phase observation value of the current epoch from the RTK observation generation module. The expected geometric phase is calculated based on the satellite geometric model and the coordinate prediction values ​​from the previous epoch. Calculate the measured phase residual Measured phase residual It reflects the combined effects of factors including the ionosphere, electromagnetic environmental disturbances, and phase rotation introduced by anti-interference weights.

[0141] S42, Based on the measured phase residual The time-varying characteristics divide the phase residual into a slowly varying bias term (representing long-term drift) and a rapidly varying drift term (representing short-term abrupt changes and transient disturbances). A Kalman filter with a long time window (10 seconds) is used to extract the slowly varying bias term, thus obtaining the slowly varying phase bias. :

[0142] ;

[0143] Further, the phase rapid shift amount was obtained. :

[0144] ;

[0145] in, Indicates the current epoch. Previous historical epoch index; This represents a long-term window set used to extract slowly varying bias terms; Indicates the historical era Measured phase residual at the location; This represents the expectation operator.

[0146] S43. Based on the anti-interference weight vector of the current epoch. Anti-interference weight vector of the previous epoch Calculate the weight increment , The weight adjustment of the spatiotemporal adaptive processing anti-interference algorithm is characterized by changes in the current interference environment.

[0147] S44. Construct an approximate linear relationship between the weight increment and the measured phase residual. ,in, For the current epoch The weight-phase sensitivity vector is used to describe the degree of influence of weight changes on the output phase; the weight-phase sensitivity vector is estimated online using the KF-EM algorithm (a hybrid algorithm combining the Kalman filter KF algorithm and the expectation-maximization EM algorithm). ,in, This represents the weight-phase sensitivity vector.

[0148] S45. Construct a sequence from the phase rapid shifts of the most recent w epochs. will sequence Input a lightweight time series forecasting model to predict the current epoch. Phase fast drift prediction :

[0149] ;

[0150] in, The lightweight time series prediction model is a time series mapping function used for short-time prediction of rapidly changing phase disturbances.

[0151] S5. Based on the phase correction reference vector, the slowly varying phase bias, and the rapidly varying phase drift prediction, a phase compensation vector is obtained by fusing these components and applying stability constraints to counteract interference from the array output signal vector. Apply phase correction to obtain the phase-corrected array output signal vector. .

[0152] Step S5 is as follows:

[0153] S51, Change the phase slowly by the bias amount As a slowly varying compensation quantity ;Predicting the phase drift of transient phase perturbations As a rapid change compensation quantity According to the weight-phase sensitivity vector Compared with the current weight increment Calculate the phase change estimate caused by the element weight perturbation in the current epoch. ;

[0154] S52, Phase correction reference vector Slow-varying compensation amount Rapidly changing compensation amount Phase change estimator Weighted fusion is performed to form the final phase compensation vector:

[0155] ;

[0156] in, The fusion weights can be adaptively adjusted based on signal quality or interference intensity.

[0157] ;

[0158] in, Indicates the first Individual elements in the historical period Phase compensation amount.

[0159] S53, anti-jamming array output signal vector By applying a complex phase rotation to each array element, the phase-corrected array element output signal is obtained. This forms the phase-corrected array output signal vector:

[0160] .

[0161] S6. Based on the phase-corrected array output signal vector RTK positioning calculation is performed, employing a double-difference observation model and introducing a dynamic scaling mechanism for observation covariance to fix ambiguity, thus obtaining the positioning result. With positioning quality indicators The observation covariance dynamic scaling mechanism adapts the covariance matrix by dynamically adjusting the weights. This enhances the robustness of the location solution under anti-interference conditions.

[0162] Step S6 is as follows:

[0163] S61. Based on the signal quality feature vector and anti-interference weight phase eigenvector and phase compensation amount Construct a weighted adaptive covariance matrix :

[0164]

[0165] in, This represents the noise covariance matrix of the observations used in the RTK positioning calculation process. When the positioning residual index is detected... When the noise covariance is large or the phase compensation is unstable, its weight is reduced by increasing the noise covariance of the corresponding observation; when the anti-interference effect is well restored and the phase compensation is stable, the noise covariance is automatically reduced to increase the contribution of effective observations to the solution. Among these, the positioning residual index... It is obtained by statistical analysis of the observation residuals or posterior position residuals output during the RTK positioning process; Indicates the satellite's elevation angle (i.e., the satellite's altitude angle); This represents the covariance scaling function.

[0166] S62. Under the double-difference observation model, using the weighted adaptive covariance matrix Phase-corrected array output signal vector By performing weighted calculations and fixing the ambiguity, the current epoch is obtained. Location results and output positioning quality indicators. ;in, Indicates the current epoch. The ambiguity fixation rate; Indicates the current epoch. Phase noise index; Indicates the current epoch. The compensation consistency index is determined by the phase compensation change between adjacent epochs. Decide; This represents the satellite's geometric parameters, determined by the geometric distribution parameters output by the receiver.

[0167] S7. Determine confidence level based on ambiguity fixation rate and phase noise index; evaluate positioning quality based on positioning quality index; and select data with confidence levels meeting set conditions as online learning sample pairs. When the positioning quality meets the set conditions, the deep learning model is trained online and its parameters are updated.

[0168] Step S7 is as follows:

[0169] S71. The high confidence criterion is: the ambiguity fixation rate satisfies... And the phase noise index meets ,in, , The thresholds for ambiguity fixation rate and phase noise index are used; if the high confidence criterion is met, the current input feature vector of the deep learning model is obtained. The final phase compensation vector The phase correction reference vector output by the deep learning model is regarded as a self-supervised signal. The sample pairs constituting online learning ;

[0170] S72. Quality evaluation criteria are: the positioning quality indicators meet the requirements. ,in, The threshold for locating quality indicators is set; if the quality evaluation criteria are met, online learning is triggered and the model parameters are updated; otherwise, online learning is not triggered and the model parameters are not updated.

[0171] This strategy is equivalent to self-imitation learning, where the model memorizes its best decision upon successful learning. The self-supervised signal construction method is also equivalent to the self-imitation learning strategy; that is, when the system localization result is determined to have high confidence, the currently effective phase compensation amount is considered as an approximately optimal decision to strengthen the model's output consistency under similar conditions.

[0172] S73, Utilizing online learning samples Define supervised learning loss :

[0173]

[0174] in, This represents the output of a deep learning model; These are the model parameters for the current epoch; this loss function gradually strengthens the mapping relationship between "feature vector → optimal compensation amount" and improves the robustness of future epochs.

[0175] The model parameters are updated using a stochastic gradient descent optimizer based on the loss gradient.

[0176]

[0177] in, For online learning rate, This represents the loss gradient.

[0178] A mini-batch strategy can be adopted to cache the most recent high-confidence samples in a queue to smooth parameter updates.

[0179] S74. Within a consecutive epoch after the deep learning model update (10 epochs in this embodiment), monitor the ambiguity fixation rate. Phase noise index Phase compensation change and positioning residual index , If any of the following occurs: ambiguity fixation rate decreases beyond the set range; phase noise index increases beyond the set range; phase compensation change exceeds the set value; or positioning residual index exceeds the set value, then the update of the model parameters is deemed to have a negative effect on the system. The model rollback mechanism is then executed to restore the model parameters to their values ​​before the update. This is used as the current optimal model parameter, and recent failed sample pairs are cleared to avoid polluting the model; if none of the above occurs, it is retained. These are the parameters of the current optimal model.

[0180] S75. For successful sample pairs that have not been cleared (successful learning), add them to the experience pool. The online learning rate is adjusted based on the current system stability. If the ambiguity fixation rate continues to increase, then the online learning rate should be increased. If the phase noise index, phase compensation change, or positioning residual index increases significantly, the online learning rate should be reduced. .

[0181] S76. Use the current optimal model parameters as the model parameters for inference in the next epoch.

[0182] Figure 2 The diagram shows the experimental results of this invention (positioning accuracy under anti-interference conditions). Under anti-interference conditions, the positioning accuracy range in the three directions of north, east, and sky (elevation) is all at the centimeter level.

[0183] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A GNSS anti-interference closed-loop phase self-calibration method based on deep learning, characterized in that, Includes the following steps: S1. Obtain the complex baseband signal of each element of the array antenna, and use the space-time adaptive processing anti-interference algorithm to obtain the anti-interference array output signal vector. With anti-interference weight vector ; where subscript Indicates an epoch, or a point in time; S2. Constructing signal quality feature vectors And extract the anti-interference weight phase feature vector. and historical positioning quality feature vector The feature vectors are concatenated to obtain the input feature vectors of the deep learning model. ; S3, input feature vector Inputting the deep learning model yields the phase correction reference vector. ; S4. Decompose the measured phase residual to obtain the phase slowly varying offset. and phase rapid drift The phase rapid change drift prediction is obtained by using the phase rapid change drift of the most recent w epochs. ; S5, Based on phase correction reference vector Phase slow-varying bias and phase rapid change drift prediction The phase compensation vector is obtained. Used to counter interference array output signal vector Apply phase correction to obtain the phase-corrected array output signal vector. ; S6, based on the phase-corrected array output signal vector Perform RTK positioning calculations to obtain the positioning results. With positioning quality indicators ; S7: Confidence is judged based on ambiguity fixation rate and phase noise index, and quality is evaluated based on positioning quality index. Data with confidence meeting the set conditions are selected as sample pairs, and the deep learning model is trained online when the positioning quality meets the set conditions.

2. The GNSS anti-interference closed-loop phase self-calibration method based on deep learning according to claim 1, characterized in that, Step S1 is as follows: S11. Receive the original radio frequency time domain signal from each element of the array antenna. The original radio frequency time domain signal contains satellite navigation signal, interference signal and noise. Perform down-conversion, bandpass filtering and gain control on the original radio frequency time domain signal of each element to shift the effective components in the working frequency band to intermediate frequency or baseband to obtain the complex intermediate frequency / complex baseband signal of each element. S12. Perform analog-to-digital conversion on the complex intermediate frequency / complex baseband signals of each array element to obtain the discrete complex baseband sampling sequence of each array element; in each epoch... The complex baseband sampled values ​​of each array element are arranged into an array data vector according to the array element index. : ； in, Indicates the first Individual elements in the historical period The complex baseband sampled values; the total number of array elements is ; S13. Within a preset time or snapshot window, based on the array data vector Estimate the interference plus noise covariance matrix covariance matrix The diagonal elements represent the received power of each array element, and the off-diagonal elements represent the spatial correlation between array elements. S14. Based on the estimated interference plus noise covariance matrix and the steering vector of the desired satellite signal An optimization problem for space-time adaptive processing is constructed, which minimizes interference and noise power while satisfying the desired satellite signal gain constraint, and the anti-interference weight vector is obtained by solving the problem. : ； ； in, Indicates the first The weight range of each array element, Indicates the first Individual elements in the historical period The weighted phase adjustment amount is the weighted phase characteristic. Indicates the epoch For the first Unit amplitude complex phase rotation applied to each array element channel, Indicates the first Individual elements in the historical period The anti-interference weight, where j represents the imaginary unit; S15. Utilizing anti-interference weight vectors For array data vectors Weighted synthesis is performed to obtain the output signal vector of the anti-interference array. : ； in, Indicates the first Individual elements in the historical period The complex baseband signal after space-time adaptive processing, i.e., the first... Individual elements in the historical period The anti-interference output signal .

3. The GNSS anti-interference closed-loop phase self-correction method based on deep learning according to claim 1, characterized in that, Step S2 is detailed below: S21, Output signal vector of the anti-jamming array Conduct signal quality assessments for each satellite. The carrier-to-noise ratio is calculated using the relevant output power and noise estimates, thus constructing a signal quality feature vector. : ； in, Indicates the first A satellite in the era The carrier-to-noise ratio; Indicates the total number of satellites; This indicates the useful signal power of the relevant output. This represents the equivalent noise power spectral density or noise estimate. S22. Extract the phase features of the anti-interference weights from the anti-interference weight vector. Extract the weighted phase feature of each array element, i.e., the weighted phase adjustment amount. This constitutes the anti-interference weighted phase eigenvector. : ； The total number of array elements is ; S23. Obtain the ambiguity fixation rate of the previous epoch from the RTK positioning solution module. and phase noise index This constitutes the historical positioning quality feature vector. : ； S24. Concatenate the three feature vectors obtained in a predetermined order to form the input feature vector of the deep learning model: ； in, This represents the concatenation operation of eigenvectors; This represents the concatenated feature vector; Normalization or standardization is performed on the concatenated feature vectors to bring all features into a uniform dimension and range, thus forming the final input feature vector. .

4. The GNSS anti-interference closed-loop phase self-correction method based on deep learning according to claim 1, characterized in that, Step S3 is detailed below: S31. Input feature vector Deep learning models input into a multilayer perceptron structure Linear mapping and activation operations are performed sequentially according to the network hierarchy: ； The output of the deep learning model is obtained through a linear mapping of the output layer. ; in, , This represents the weight matrix and bias vector of the i-th layer; This represents the hidden layer output; the deep learning model has a total of s layers; This represents the linear rectification function, also known as the activation function. Indicates the current epoch. Model parameters; S32, the output of the deep learning model is The phase correction reference quantity of each array element is used to obtain the phase correction reference vector. ;in, The first inference obtained from the model represents the... Phase correction reference quantity for each array element; S33. To ensure that the output of the deep learning model does not compromise the stability of subsequent phase compensation, an optional consistency check is performed: First, an amplitude limit is applied to the amplitude, with the following condition: , The first step is to set an amplitude threshold; the second step is to apply stability constraints to changes in adjacent epochs, with the following constraints: , This is the threshold for the amount of change; if the constraint is triggered, the output of the deep learning model is pruned to obtain the final phase correction reference vector. .

5. The GNSS anti-interference closed-loop phase self-calibration method based on deep learning according to claim 1, characterized in that, Step S4 is as follows: S41. Extract the measured carrier phase observation value of the current epoch from the RTK observation generation module. The expected geometric phase is calculated based on the satellite geometric model and the coordinate prediction values ​​from the previous epoch. Calculate the measured phase residual ; S42, Based on the measured phase residual The time-varying characteristics of the phase residual divide it into a slowly varying bias term and a rapidly varying drift term. A Kalman filter is used to extract the slowly varying bias term to obtain the slowly varying phase bias. : ； Thus, the phase rapid shift amount is obtained. : ； in, Indicates the current epoch. Previous historical epoch index; The set of windows representing the Kalman filter; Indicates the historical era Measured phase residual at the location; Represents the expectation operator; S43. Based on the anti-interference weight vector of the current epoch. Anti-interference weight vector of the previous epoch Calculate the weight increment , Characterizes the weight adjustment of the spatiotemporal adaptive processing anti-interference algorithm due to changes in the current interference environment; S44. Construct an approximate linear relationship between the weight increment and the measured phase residual. ,in, For the current epoch The weight-phase sensitivity vector is used to describe the degree of influence of weight changes on the output phase. A hybrid algorithm combining the Kalman filter algorithm and the expectation-maximization algorithm is used to estimate the weight-phase sensitivity vector online. ; Represents the weight-phase sensitivity vector; S45. Construct a sequence from the phase rapid shifts of the most recent w epochs. will sequence Input a time series prediction model to predict the current epoch. Phase fast drift prediction : ； in, This represents a time series prediction model.

6. The GNSS anti-interference closed-loop phase self-calibration method based on deep learning according to claim 5, characterized in that, Step S5 is as follows: S51, Change the phase slowly by the bias amount As a slowly varying compensation quantity ; Predicting phase fast drift As a rapid change compensation quantity According to the weight-phase sensitivity vector Compared with the current weight increment Calculate the phase change estimate ; S52, Phase correction reference vector Slow-varying compensation amount Rapidly changing compensation amount Phase change estimator Weighted fusion is performed to form the final phase compensation vector: ； in, For weighting; ； in, Indicates the first Individual elements in the historical period Phase compensation amount; total number of array elements is ; S53, anti-jamming array output signal vector By applying a complex phase rotation to each array element, the phase-corrected array element output signal is obtained. This forms the phase-corrected array output signal vector: ； in, Represents the output signal vector of the anti-interference array The Middle The anti-interference output signal of each array element.

7. The GNSS anti-interference closed-loop phase self-correction method based on deep learning according to claim 6, characterized in that, In step S6, a double-difference observation model is used in the RTK positioning solution process, and a dynamic scaling mechanism for observation covariance is introduced to fix ambiguity, as detailed below: S61. Based on the signal quality feature vector and anti-interference weight phase eigenvector and phase compensation amount Construct a weighted adaptive covariance matrix ; S62. In the double-difference observation model, the weighted adaptive covariance matrix is ​​used. Phase-corrected array output signal vector Weighted calculations are performed, and combined with fixed ambiguity, to obtain the positioning result. Output quality indicators ;in, Indicates the current epoch. The ambiguity fixation rate; Indicates the current epoch. Phase noise index; Indicates the current epoch. The compensation consistency index is determined by the phase compensation change between adjacent epochs. Decide; This represents the satellite's geometric parameters, determined by the geometric distribution parameters output by the receiver.

8. The GNSS anti-interference closed-loop phase self-calibration method based on deep learning according to claim 7, characterized in that, Step S7 is detailed below: S71. The high confidence criterion is: the ambiguity fixation rate satisfies... And the phase noise index meets ,in, , The thresholds for ambiguity fixation rate and phase noise index are used; if the high confidence criterion is met, then a sample pair for online learning is formed. ; S72. Quality evaluation criteria are: the positioning quality indicators meet the requirements. ,in, The threshold for locating quality indicators is set; if the quality evaluation criteria are met, online learning is triggered and the model parameters are updated; otherwise, online learning is not triggered and the model parameters are not updated. S73, using online learning samples to... Define supervised learning loss : in, This represents the output of a deep learning model; These are the model parameters for the current epoch; The stochastic gradient descent optimizer is used to update the model parameters based on the loss gradient. in, For online learning rate, Represents the loss gradient; S74. Monitor the ambiguity fixation rate within a consecutive epoch after the deep learning model is updated. Phase noise index Phase compensation change And positioning residual index, If any of the following occurs: ambiguity fixation rate decreases beyond the set range; phase noise index increases beyond the set range; phase compensation change exceeds the set value; or positioning residual index exceeds the set value, then the update of the model parameters is deemed to have a negative effect. The model rollback mechanism is then executed to restore the model parameters to their values ​​before the update. Use these as the current optimal model parameters and clear the sample pairs from the most recent period, i.e., failed sample pairs; otherwise, retain them. As the current optimal model parameters; S75. For successful sample pairs that have not been cleared, add them to the experience pool; S76. Use the current optimal model parameters as the model parameters for the next epoch.

9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a deep learning-based GNSS anti-interference closed-loop phase self-calibration method as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the deep learning-based GNSS anti-interference closed-loop phase self-calibration method as described in any one of claims 1 to 8.

Images