Kalman neural network-based low earth orbit satellite signal frequency estimation method and system

Abstract

The application discloses a kind of low-orbit satellite signal frequency estimation method and system based on Kalman neural network, mainly solve the problem of low frequency estimation precision and poor robustness in non-cooperative high dynamic scene in prior art.The scheme is: initialization parameter, generate carrier wipe-off signal to wipe off received signal carrier, calculate the remaining frequency estimation value, and obtain the measurement value by observation;Build an end-to-end network model including high dynamic feature extraction unit, two-layer gated recurrent unit, self-attention fusion unit and constrained output unit;The network model is supervised and trained using the measurement value and state estimation increment;The measurement value and state estimation increment are input into the trained network, and the Kalman gain prediction value is output, the state posteriori estimation value at the current time is calculated as the input for the next time cycle update, to realize continuous online high-precision estimation of satellite signal frequency.The application can realize real-time, robust and high-precision estimation of low-orbit satellite downlink signal frequency, and can be used for non-cooperative satellite navigation and positioning.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, and specifically relates to a method and system for estimating the frequency of low-Earth orbit satellite signals, which can be used in non-cooperative satellite navigation and positioning systems. Background Technology

[0002] In non-cooperative, high-dynamic application scenarios, rapid, stable, and high-precision estimation of the signal carrier frequency is a crucial fundamental parameter for realizing subsequent functions such as opportunistic signal navigation, target localization, and electromagnetic situational awareness. However, under non-cooperative reception conditions, the signal modulation structure, symbol format, and time-frequency parameters are usually unknown. Furthermore, it is difficult to obtain accurate state transition models F and the statistical characteristics of system noise Q and measurement noise R. Traditional Kalman filtering methods, which rely on precise dynamics and noise priors, struggle to maintain stable convergence and high-precision estimation, thus limiting their engineering applicability in non-cooperative, high-dynamic scenarios.

[0003] Patent application CN201310631478.X discloses a "carrier tracking method and loop for a GPS signal receiver." The implementation steps are as follows: receiving GPS satellite signals through an antenna, performing down-conversion and correlation processing; inputting the correlation results into a discriminator to calculate the frequency difference or phase difference between the received signal and the local signal; wherein, when the carrier loop is a frequency-locked loop, the discriminator is a frequency discriminator; when the carrier loop is a phase-locked loop, the discriminator is a phase discriminator; filtering the discrimination result using a Kalman filter; inputting the updated output result into a loop filter for further filtering, thereby outputting the carrier Doppler frequency, and generating a local carrier based on the carrier Doppler frequency and the carrier frequency. This method suffers from drawbacks because the traditional Kalman filter relies on an accurate system model and prior statistical characteristics of noise. Therefore, in complex scenarios with low signal-to-noise ratio, strong dynamics, and insufficient prior information, model mismatch and noise non-stationarity can easily lead to filter divergence and decreased tracking accuracy. Furthermore, it is difficult to adaptively match the dynamic characteristics of the signal, limiting robustness and generalization ability.

[0004] Patent document CN201610681535.9 discloses "A tracking loop for a satellite navigation receiver based on Kalman filtering." The implementation scheme is as follows: the intermediate frequency signal after down-conversion sampling by the receiver is processed by a mixing module and a correlation module to obtain a baseband signal, which is then sent to a coherent integration module. The coherent integration result output by the coherent integration module is sent to a code phase detector, a carrier frequency detector, and a carrier phase detector, respectively. A Kalman filter filters the aforementioned phase and frequency detection results, and finally uses the filtered estimates to control the carrier numerically controlled oscillator and the code numerically controlled oscillator to generate updated local carrier and code signals, completing the closed-loop tracking. This method suffers from drawbacks because the traditional Kalman filter used is highly dependent on the system state model and noise statistics. Therefore, in real-world scenarios with low signal-to-noise ratios, high dynamic changes, and limited prior information, it is prone to model mismatch, decreased filtering accuracy, and even divergence. Furthermore, it struggles to adapt to complex time-varying channel environments, resulting in significant limitations in dynamic tracking performance and robustness. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by proposing a method and system for estimating the frequency of low-Earth orbit satellite signals based on a Kalman neural network, so as to achieve real-time, robust, and high-precision estimation of the downlink signal frequency of low-Earth orbit satellites in a highly dynamic and time-varying Doppler environment with unknown statistical characteristics.

[0006] The technical approach to achieve the objective of this invention is as follows: while retaining the recursive closed-loop framework of Kalman filtering, the state transition and observation model are embedded into the network as physical constraints, and the time-series related information in the observation sequence is deeply modeled using gated recurrent units. A self-attention mechanism that satisfies online causal constraints is used to complete the adaptive weighted fusion of historical information, and the optimal Kalman gain is adaptively learned, thereby improving the robustness of frequency estimation under conditions of unknown noise statistics and strong dynamics.

[0007] Based on the above ideas, the technical solution of the present invention includes:

[0008] 1. A method for estimating the frequency of low-Earth orbit satellite signals based on a Kalman neural network, characterized in that it includes:

[0009] (1) Initialize the various model parameters in the Kalman neural network;

[0010] (2) Calculate the state prior estimate at the current time based on the initial value or the state posterior estimate at the previous time, and calculate the state estimate increment based on the state posterior estimate and the state prior estimate at the previous time.

[0011] (3) Generate the corresponding local carrier reference signal based on the prior state estimate at the current moment;

[0012] (4) Mix the local carrier reference signal with the down-converted signal, calculate the remaining frequency estimate based on the mixed signal and the local standard signal, and then observe the frequency estimate to obtain the measured value.

[0013] (5) Construct an end-to-end network model including a high-dynamic feature extraction unit, a two-layer gated recurrent unit, a self-attention fusion unit, and a constrained output unit;

[0014] (6) Input the above measured values ​​and state estimation increments into the end-to-end network model, and train it using supervised learning to obtain a trained end-to-end network model;

[0015] (7) Input the above measured values ​​and state estimation increments into the trained network model, output the Kalman gain prediction value, and calculate the state posterior estimate value at the current time based on the prediction value, the current state prior estimate value and the measured value;

[0016] (8) The state posterior estimate is used as the input for the next moment. The above process is repeated until the set number of cycles is reached to complete the continuous online high-precision estimation of the satellite signal frequency.

[0017] 2. A low-Earth orbit satellite signal frequency estimation system based on a Kalman neural network, characterized in that it comprises:

[0018] The initialization configuration module is used to initialize the core model parameters of the Kalman neural network, including the state transition matrix and the observation matrix, as well as the initial state values, including the state prior estimate and the state posterior estimate. It also sets the maximum number of iterations for satellite signal frequency estimation.

[0019] The input feature generation module is used to calculate the state prior estimate at the current time based on the state posterior estimate at the previous time, generate a local carrier reference signal, mix it with the down-conversion signal, and combine it with the local standard signal to calculate the remaining frequency estimate to obtain the measurement value.

[0020] The network construction module builds an end-to-end network model that includes a high-dynamic feature extraction unit, a two-layer gated recurrent unit, a self-attention fusion unit, and a constrained output unit.

[0021] The network training module is used to take the state estimation increment and measurement value output by the input feature generation module as the model input, and complete the training and validation of the end-to-end network model in a supervised learning manner to obtain a converged and high-performance Kalman gain prediction model.

[0022] The network testing module uses the state estimation increment and measurement value output from the input feature generation module as inputs to the trained network model to predict the Kalman gain.

[0023] The output result generation module is used to calculate the posterior estimate of the current state based on the predicted Kalman gain, the prior estimate of the current state, and the measured value. This process is repeated until the termination condition is met, and finally, the continuous online high-precision satellite signal frequency estimation result is output.

[0024] Compared with the prior art, the present invention has the following advantages:

[0025] Firstly, this invention is based on an adaptive estimation mechanism driven by data and constrained by physical models. By jointly modeling the dynamic characteristics of the system and the statistical features of observations through gated recurrent units and self-attention fusion networks, it achieves dynamic screening and weighted representation of key information. It can adaptively adjust the filter gain and state update process without the need for precise prior model parameters, effectively reducing the impact of model mismatch and noise uncertainty on frequency estimation performance, and significantly improving estimation stability and robustness.

[0026] Secondly, this invention effectively characterizes the nonlinear and non-stationary characteristics in the frequency evolution process based on a gated cyclic unit with time-series modeling capabilities, and combines a finite window self-attention mechanism to adaptively weight and fuse key historical information, thereby achieving rapid convergence in the dynamic change stage and high-precision frequency estimation in the stable stage, effectively improving the overall estimation performance and system adaptability. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the implementation of the low-Earth orbit satellite signal frequency estimation method based on a Kalman neural network, as described in this invention.

[0028] Figure 2 This is a schematic diagram of the end-to-end network model structure constructed by the method of the present invention;

[0029] Figure 3 This is a block diagram of the low-Earth orbit satellite signal frequency estimation system based on Kalman neural network of the present invention;

[0030] Figure 4 This invention provides a robustness analysis of the model under observation noise mismatch.

[0031] Figure 5 This invention provides a robustness analysis of the model under process noise mismatch. Detailed Implementation

[0032] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0033] Example 1: A method for estimating the frequency of low-Earth orbit satellite signals based on a Kalman neural network.

[0034] Reference Figure 1 The implementation steps of this example include the following:

[0035] Step 1: Initialize the various model parameters and signal states of the Kalman neural network.

[0036] set up It is the first The carrier phase state quantity at each sub-accumulation point, where For signal phase, The signal Doppler frequency, The Doppler change rate of the signal;

[0037] Let the initial value of the Kalman filter be... ,in For the initial phase, The initial Doppler frequency, The initial Doppler rate of change;

[0038] Initialize the state transition matrix ,in The sampling period;

[0039] Initialize the observation matrix .

[0040] Step 2: Generate a local reference signal based on the prior estimate of the current state and perform carrier erasure on the received signal.

[0041] 2.1) Constructing the received signal model:

[0042] 2.1.1) Let the signal transmitted by the transmitter be... ,in It is a deterministic repetitive signal with a period of . , It is user-side modulation information, usually random data. and They are unrelated.

[0043] 2.1.2) Transmit signal After propagation through an additive white Gaussian noise channel, the intermediate frequency received signal is obtained. :

[0044] ,

[0045] In the formula, Indicates the apparent delay between the transmitted and received signals. Indicates the receiving frequency. This represents zero-mean complex Gaussian white noise;

[0046] 2.1.3) For intermediate frequency received signals The received signal is obtained by mixing and filtering. :

[0047] ,

[0048] In the formula, It consists of user data and channel noise. To receive signals The phase;

[0049] 2.1.4) The received signal after mixing and filtering Sampling is performed to obtain discrete received signals. :

[0050] ,

[0051] In the formula, yes The discrete-time corresponding sequence, period , The sampling period; and They are the first Discrete-time carrier phase and code phase of the received signal at each sub-accumulator; yes The discrete-time corresponding sequence;

[0052] 2.2) Generate carrier erase signal:

[0053] 2.2.1) For The posterior estimate of the state at time 1 By performing recursion, we obtain State prior estimate at time 1 :

[0054] ,

[0055] 2.2.2) For prior state variables The carrier phase of the carrier erase signal is reconstructed. :

[0056] ,

[0057] 2.2.3) Based on reconstructed carrier phase Reconstructing the local carrier erase signal :

[0058] ,

[0059] In the formula, The signal amplitude;

[0060] 2.3) Regarding the received signal Perform carrier erasure to obtain the signal after carrier erasure. :

[0061] ,

[0062] In the formula, This refers to the signal state quantity after carrier erasure. For user data and channel noise.

[0063] Step 3: Perform cross-correlation calculation between the carrier-erased signal and the local standard signal to calculate the estimated remaining frequency, and then observe the estimated frequency to obtain the measured value.

[0064] 3.1) Generate local standard signals ;

[0065] Because the blind receiver does not transmit deterministic repetitive signals in the signal. Prior knowledge, from the initial estimate Initially, each moment is based on the received signal. Deterministic repetitive signals in Update local standard signal ;

[0066] 3.2) The signal after carrier erasure The frequency domain signal is obtained by performing a Fourier transform. ;

[0067] 3.3) The signal after carrier erasure With local standard signals Cross-correlation calculations were performed to obtain the estimated residual frequency. :

[0068] ,

[0069] In the formula, Signal state quantity after carrier erasure The second state component in;

[0070] 3.4) Signal state variables after carrier erasure Observations were conducted to obtain measured values. :

[0071]

[0072] in, This is the observation matrix.

[0073] Step 4: Construct input features based on the observations and state estimation increments.

[0074] 4.1) Based on the prior state estimate at the current moment and posterior state estimates Calculate the state estimation increment :

[0075] ,

[0076] 4.2) Utilization Observations of time and State estimation increment at time step ,constitute Input features at time step :

[0077] .

[0078] Step 5: Build an end-to-end network model.

[0079] Reference Figure 2 The implementation of this step includes the following:

[0080] 5.1) Establish a high-dynamic feature extraction unit comprising multiple fully connected layers to perform nonlinear mapping on input features and extract higher-order features related to the dynamic changes in signal state. :

[0081] ,

[0082] in, , For learnable parameters, It is a non-linear activation function. This represents the input feature vector. Estimate the increment for the state. These are measured values;

[0083] 5.2) Establish a two-layer gated recurrent unit (GRU) in each layer, consisting of a reset gate, an update gate, and candidate hidden states. The reset gate filters the previous hidden state and participates in generating candidate hidden states by linearly transforming and activating the current input and the previous hidden state. The update gate controls the fusion ratio of the previous hidden state and candidate hidden states in the final hidden state through independent linear transformation and activation. The two work together to selectively memorize and update sequence information. That is, this first-layer gated recurrent unit is used to capture the first... Short-term rapid dynamic features of hidden states at time points The second-level gated loop unit is used to capture the first... Hidden state of medium- and long-term statistical changes at any given time :

[0084] ,

[0085] ,

[0086] in, Indicates the first loop unit in The hidden state at any given moment. Indicates the second-level loop unit in The hidden state at any given moment;

[0087] 5.3) Establish a self-attention fusion unit employing a self-attention mechanism to perform weighted fusion of historical information and generate contextual feature representations to characterize temporal contextual relationships. :

[0088] ,

[0089] in, To calculate attention weights, , , , The weight matrix is ​​trainable. , The weight matrix is ​​trainable.

[0090] 5.4) Construct constrained output units comprising multiple fully connected layers to fuse attention features. Mapped to Kalman gain estimate :

[0091] ,

[0092] in, and For trainable weight parameters, This represents the Sigmoid activation function, which is used to constrain the output result to a preset interval [0,1] to meet the physical feasibility constraint of Kalman gain, and to apply L2 regularization constraint to the trainable parameters of the output network to improve generalization stability.

[0093] 5.5) The high dynamic feature extraction unit, the two-layer gated recurrent unit, the self-attention fusion unit and the constrained output unit are cascaded in sequence to form an end-to-end network model.

[0094] In this example, but not limited to the high-dynamic feature extraction unit, the number of fully connected layers is three, wherein the first fully connected layer is used to process the input feature vector. The system performs dimensional mapping and initial nonlinear transformation. The second fully connected layer acts as a hidden layer to further extract higher-order features. The third fully connected layer is used to map the hidden layer features to higher-order features of the target dimension. The constrained output unit has two fully connected layers, where the first fully connected layer uses a nonlinear activation function to... Intermediate features are extracted, and the second fully connected layer maps the intermediate features to the Kalman gain dimension and constrains them to the [0,1] interval using Sigmoid.

[0095] Step 6: Train the end-to-end network model using supervised learning.

[0096] 6.1) Set the training rounds to 200 and the training batches to 32;

[0097] 6.2) Use the Adam optimizer, with an initial learning rate of 0.001 and a weight decay factor of 0.0001;

[0098] 6.3) Input the measured values ​​and state estimation increments into the end-to-end network for forward propagation, and calculate the Kalman gain prediction output. ;

[0099] 6.4) Calculate the posterior state estimate based on the predicted Kalman gain. Use it with the true value of the state. Construct the loss function and calculate the loss gradient of the signal state. :

[0100] ;

[0101] 6.5) The gradient is backpropagated along the four cascaded units in the network—the high dynamic feature extraction unit, the two-layer gated recurrent unit, the self-attention fusion unit, and the constrained output unit—using the chain rule to update the network parameters of each unit in turn.

[0102] 6.6) Repeat steps 6.3) to 6.5) until the maximum number of iterations is reached, and the trained end-to-end network model is obtained.

[0103] Step 7: Use the trained end-to-end network model to obtain high-precision satellite signal frequency estimation results.

[0104] 7.1) Input feature vector Input the trained network model to obtain the Kalman gain prediction value. ;

[0105] 7.2) Based on Kalman gain prediction Prior state estimation and measured values Calculate the posterior state estimate at the current time. :

[0106] ,

[0107] 7.3) The posterior state estimate As input for the next time step, and 7.1) and 7.2) are executed repeatedly until the termination condition is met, the final output is a continuous online high-precision satellite signal frequency estimation result.

[0108] Example 2: Low-Earth Orbit Satellite Signal Frequency Estimation System Based on Kalman Neural Network

[0109] Reference Figure 3 This example includes: initialization configuration module 1, input feature generation module 2, network construction module 3, network training module 4, network testing module 5, and output result generation module 6. Among them, network construction module 3 includes: high dynamic feature extraction submodule 31, two-layer gated recurrent unit submodule 32, self-attention fusion submodule 33, and self-attention fusion submodule 34.

[0110] The working principle of the entire system is as follows:

[0111] The initialization configuration module 1 is used to initialize the core model parameters of the Kalman neural network, including the state transition matrix and the observation matrix, as well as the initial state values, including the state prior estimate and the state posterior estimate, and to pass the model parameters and the initial state values ​​into the input feature generation module 2.

[0112] The input feature generation module 2 is used to calculate the initial state value and model parameters passed in by the initialization configuration module 1 to obtain the state prior estimate value at the current time, generate the local carrier reference signal, mix it with the down-conversion signal, and calculate the remaining frequency estimate value in combination with the local standard signal to obtain the measurement value, and pass the input features into the network training module 4 and the network testing module 5.

[0113] The network construction module 3 is used to construct an end-to-end network model consisting of a high-dynamic feature extraction submodule 31, a two-layer gated recurrent unit submodule 32, a self-attention fusion submodule 33, and a self-attention fusion submodule 33.

[0114] The high-dynamic feature extraction submodule 31 is used to perform nonlinear transformation and mapping on the input features, adaptively mine high-order features related to the high-dynamic characteristics of the system, and pass the high-order features into the two-layer gated loop unit submodule 32.

[0115] The two-layer gated loop unit submodule 32 is used to perform serialized recursive modeling on the high-order features input from the high-dynamic feature extraction submodule 31 to obtain the hidden state at the current time, and input the hidden state at the current time and the most recent L historical hidden states into the self-attention fusion submodule 33.

[0116] The self-attention fusion submodule 33 is used to perform attention mechanism weighted fusion of the current hidden state and the most recent L historical hidden states passed in by the two-layer gated loop unit submodule 32 to generate fusion features that can accurately characterize the global temporal context association, and then pass the context features into the constrained output submodule 34.

[0117] The constrained output submodule 34 is used to map the context features passed from the self-attention fusion submodule 33 to the legal value space, and finally outputs the Kalman gain estimate that satisfies the constraints.

[0118] The network training module 4 takes the state estimation increment and measurement value passed from the input feature generation module 2 as the model input, and completes the training and verification of the end-to-end network model using supervised learning, so as to obtain a converged and high-performance Kalman gain prediction model.

[0119] The network testing module 5 uses the state estimation increment and measurement value passed from the input feature generation module 2 to input the network model trained by the network training module 4, predicts the Kalman gain, and then passes the Kalman gain to the output result generation module 6.

[0120] The output result generation module 6 is used to calculate the posterior estimate of the current state based on the Kalman gain predicted by the network test module 5, the prior estimate of the current state, and the measured value, and to repeatedly execute this process until the termination condition is met, and finally output the continuous online high-precision satellite signal frequency estimation result.

[0121] It should be noted that the above functional modules can be implemented, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, they can be implemented, in whole or in part, as program instruction products. A program instruction product includes one or a set of program instructions. When the program instructions are loaded and executed on a computer, the described process or function is generated, in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The program instructions can be stored in a computer-readable and writable storage medium, or transferred from one computer's readable and writable storage medium to another.

[0122] In this embodiment, the direct coupling or communication connection between the modules can be achieved through indirect coupling or communication connection via interfaces, devices, or modules. The functional modules in this embodiment can dynamically reside within a single processing unit, or each module can exist physically independently, or two or more modules can dynamically reside within a single processing unit. When these dynamic components are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable and writable storage medium. This storage medium can be a memory, disk, or optical disc, etc.

[0123] The effects of this invention will be further illustrated below with simulation experiments:

[0124] 1. Simulation experimental conditions:

[0125] The hardware platform for the simulation experiment is: i9-14900K CPU processor, 64GB memory, and GetForce RTX 4090 (22GB) graphics card.

[0126] The software platform for the simulation experiment is: Windows 10 operating system, Python 3.8 and PyTorch 12.0.

[0127] The input data used in the simulation experiment was simulated Starlink satellite downlink service signals, and the process noise covariance was used when generating the signal state. and observation noise covariance The signal is continuously observed and sampled, with a sampling time interval. Kalman filtering process noise covariance and observation noise covariance .

[0128] 2. Simulation Experiment Content

[0129] Simulation Experiment 1: Under the above conditions, the observation noise was compared using the methods of this invention and the traditional Kalman filter. Frequency estimation of low-Earth orbit satellite signals under mismatch conditions, the results are as follows: Figure 4 As shown.

[0130] from Figure 4 It is evident that, under observation noise mismatch conditions, the frequency estimation accuracy of the method described in this invention is significantly superior to that of the traditional Kalman filtering method. This is because the traditional Kalman filtering is highly sensitive to the accuracy of the observation noise parameters: when the observation noise is overestimated, the filter assigns too low a weight to the observation information during the update phase, resulting in a slower system response and difficulty in timely tracking changes in the target state; conversely, when the observation noise is underestimated, the filter relies excessively on noisy observations, introducing additional measurement noise interference. Both of these mismatch scenarios lead to a root mean square error significantly higher than that of the Kalman filtering under ideal matching conditions. In contrast, this invention, without prior knowledge of the true statistical characteristics of the observation noise, can maintain stable estimation accuracy across different time steps, and its overall error level is significantly lower than that of Kalman filtering under various observation noise mismatch conditions, indicating that this invention possesses stronger robustness against noise mismatch. This result demonstrates that this invention can learn an effective trade-off between predictive and observation information through a data-driven approach, thereby mitigating performance degradation under conditions of inaccurate observation noise statistics.

[0131] Simulation Experiment 2: Under the above conditions, the process noise was compared using the methods of this invention and the traditional Kalman filter. Frequency estimation of low-Earth orbit satellite signals under mismatch conditions, the results are as follows: Figure 5 As shown.

[0132] from Figure 5 It is evident that, under process noise mismatch conditions, the frequency estimation accuracy of the method described in this invention is significantly superior to that of the traditional Kalman filter method. This is because the traditional Kalman filter is highly sensitive to the accuracy of the observed noise parameters: when the process noise is underestimated, the traditional Kalman filter incorrectly assumes that the target motion is relatively smooth, leading to a significant decrease in the filter's dynamic response capability and its inability to adapt to strong maneuvering behavior, thus generating significant hysteresis errors; when the process noise is overestimated, although the Kalman filter maintains the response speed to some extent, its estimation results exhibit significant jitter due to the introduction of too many process uncertainty assumptions, and the overall accuracy is still lower than that under ideal matching conditions. In contrast, this invention can maintain a low error level close to that of the ideal Kalman filter even under conditions lacking accurate prior information on process noise and where the target motion exhibits strong nonlinearity and high maneuverability. This indicates that the end-to-end network structure constructed in this invention can not only effectively capture the implicit dynamic characteristics of the target motion, but also adaptively adjust the filter gain in complex environments where process noise is unknown or time-varying, thereby achieving near-theoretical optimal estimation performance.

[0133] The above description is merely a specific example of the present invention and does not constitute any limitation on the present invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and details without departing from the principles and structure of the present invention. For example, the end-to-end network structure in Example 1 of the present invention is not limited to the use of gated recurrent units and self-attention mechanisms. Other network structures suitable for time-series signal processing are also applicable to the framework of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A method for estimating the frequency of low-Earth orbit satellite signals based on a Kalman neural network, characterized in that, include: (1) Initialize the various model parameters in the Kalman neural network; (2) Calculate the state prior estimate at the current time based on the initial value or the state posterior estimate at the previous time, and calculate the state estimate increment based on the state posterior estimate and the state prior estimate at the previous time. (3) Generate the corresponding local carrier reference signal based on the prior state estimate at the current moment; (4) Mix the local carrier reference signal with the down-converted signal, calculate the remaining frequency estimate based on the mixed signal and the local standard signal, and then observe the frequency estimate to obtain the measured value. (5) Construct an end-to-end network model including a high-dynamic feature extraction unit, a two-layer gated recurrent unit, a self-attention fusion unit, and a constrained output unit; (6) Input the above measured values ​​and state estimation increments into the end-to-end network model, and train it using supervised learning to obtain a trained end-to-end network model; (7) Input the above measured values ​​and state estimation increments into the trained network model, output the Kalman gain prediction value, and calculate the state posterior estimate value at the current time based on the prediction value, the current state prior estimate value and the measured value; (8) The state posterior estimate is used as the input for the next moment. The above process is repeated until the set number of cycles is reached to complete the continuous online high-precision estimation of the satellite signal frequency.

2. The method according to claim 1, characterized in that, In step (1), the initialization of the model parameters in the Kalman neural network is to set the initial state value as follows: The state transition matrix is ​​F, and the observation matrix is ​​H, where For the initial phase, The initial Doppler frequency, The initial Doppler rate of change.

3. The method according to claim 1, characterized in that: In step (2), the prior state estimate for the current moment is calculated based on the initial value or the posterior state estimate from the previous moment, using the following formula: ； in, This is the posterior estimate of the state at time k-1. The prior state estimate at time k; In step (2), the state estimation increment is calculated based on the posterior state estimate and the prior state estimate from the previous time step. The formula is as follows: ； in, Let be the posterior estimate of the state at time k.

4. The method according to claim 1, characterized in that, The generation of the corresponding local carrier reference signal based on the prior state estimate at the current time in (3) includes the following implementation: (3a) Estimated by prior state values Reconstructing the carrier phase of the local carrier reference signal : ； in, The sampling period is This represents the number of sampling points; (3b) Based on reconstructed carrier phase Reconstructing the local carrier reference signal : ， in, This represents the signal amplitude.

5. The method according to claim 1, characterized in that, In step (4), the remaining frequency estimate is calculated, and then the frequency estimate is observed to obtain the measured value. The calculation process is as follows: (4a) The signal after carrier erasure The frequency domain signal is obtained by performing a Fourier transform. The remaining frequency estimate is obtained by cross-correlation calculation with the local standard signal. : ， in, It is a local standard signal. Signal state quantity after carrier erasure The second state component in; (4b) Signal state variables after carrier erasure Observations were conducted to obtain measured values. : ， in, It is the observation matrix.

6. The method according to claim 1, characterized in that, The end-to-end network model constructed in (5) includes a high-dynamic feature extraction unit, a two-layer gated recurrent unit, a self-attention fusion unit, and a constrained output unit, comprising: (5a) Establish a high dynamic feature extraction unit composed of multiple fully connected layers to perform nonlinear mapping on the input features and extract high-order features related to the dynamic changes of the system. : ； in, , For learnable parameters, It is a non-linear activation function. This represents the input feature vector. Estimate the increment for the state. These are measured values; (5b) Establish a two-layer gated loop unit consisting of a reset gate, an update gate, and candidate hidden states, which are used to capture the first hidden state, respectively. Short-term rapid dynamic features of hidden states at time points and the latent state of medium- and long-term statistical changes : ， ， in, Indicates the first loop unit in The hidden state at any given moment. Indicates the second-level loop unit in The hidden state at any given moment; (5c) Establish a self-attention fusion unit composed of self-attention mechanisms to perform weighted fusion of historical information and generate contextual feature representations to characterize temporal contextual relevance. : ， in, To calculate attention weights, , , , The weight matrix is ​​trainable. , The weight matrix is ​​trainable. (5d) Construct a constrained output unit consisting of multiple fully connected layers to fuse attention features. Mapped to Kalman gain estimate : ， in, and For trainable weight parameters, This represents the Sigmoid activation function, which is used to constrain the output result to a preset interval [0,1] to meet the physical feasibility constraint of Kalman gain, and to apply L2 regularization constraint to the trainable parameters of the output network to improve generalization stability. (5e) The above-mentioned high dynamic feature extraction unit, two-layer gated recurrent unit, self-attention fusion unit and constrained output unit are cascaded in sequence to form an end-to-end network model.

7. The method according to claim 1, characterized in that, The supervised learning method used in (6) to train the end-to-end network model includes: (6a) The number of training rounds is set to 200, and the number of training batches is set to 32; (6b) Use the Adam optimizer with an initial learning rate of 0.001 and a weight decay factor of 0.0001; (6c) Forward propagation is performed on the measurement values ​​and state estimation increments input into the network to calculate the Kalman gain prediction output; (6d) The posterior state estimate is calculated based on the predicted Kalman gain. Use it with the actual state value Construct the loss function and calculate the loss gradient of the signal state. : ， (6e) The gradient is backpropagated along the four cascaded units using the chain rule, and the network parameters of each unit are updated sequentially. (6f) Repeat (6c) to (6e) until the maximum number of iterations is reached to obtain the trained end-to-end network model.

8. The method according to claim 1, characterized in that, In step (7), the posterior state estimate at the current moment is calculated based on the Kalman gain prediction, the prior state estimate at the current moment, and the measured state. : ； in, This is the prior estimate of the state at the current moment. This is the predicted Kalman gain value. These are measured values.

9. A low-Earth orbit satellite signal frequency estimation system based on a Kalman neural network, characterized in that, include: The initialization configuration module is used to initialize the core model parameters of the Kalman neural network, including the state transition matrix and the observation matrix, as well as the initial state values, including the state prior estimate and the state posterior estimate. At the same time, it sets the maximum number of iterations for satellite signal frequency estimation. The input feature generation module is used to calculate the state prior estimate at the current time based on the state posterior estimate at the previous time, generate a local carrier reference signal, mix it with the down-conversion signal, and combine it with the local standard signal to calculate the remaining frequency estimate to obtain the measurement value. The network construction module builds an end-to-end network model that includes a high-dynamic feature extraction unit, a two-layer gated recurrent unit, a self-attention fusion unit, and a constrained output unit. The network training module is used to take the state estimation increment and measurement value output by the input feature generation module as the model input, and complete the training and validation of the end-to-end network model in a supervised learning manner to obtain a converged and high-performance Kalman gain prediction model. The network testing module uses the state estimation increment and measurement value output from the input feature generation module as inputs to the trained network model to predict the Kalman gain. The output result generation module is used to calculate the posterior estimate of the current state based on the predicted Kalman gain, the prior estimate of the current state, and the measured value. This process is repeated until the termination condition is met, and finally, the continuous online high-precision satellite signal frequency estimation result is output.

10. The system according to claim 9, characterized in that, The end-to-end network model includes: The high-dynamic feature extraction submodule is used to perform nonlinear transformations and mappings on the input features, adaptively mine high-order feature representations related to the high-dynamic characteristics of the system, and enhance the model's ability to represent complex dynamic changes. Two-layer gated recurrent unit submodules are used to perform serialized recursive modeling of temporal features, effectively capturing the dependencies and evolution patterns in the time dimension, and improving the long-term modeling capability of dynamic time-series signals; The self-attention fusion submodule is used to perform weighted fusion of historical temporal information based on the current hidden state and the L most recent historical hidden states through an attention mechanism, so as to generate fusion features that can accurately characterize the global temporal context. The constrained output submodule is used to map the attention-fused features to the legal value space and finally output the Kalman gain estimate that satisfies the constraints, ensuring the stability and rationality of the filtering update process.

Images