Intelligent Identification and Adaptive Acquisition Method and System for Multi-Constellation Satellite Signals

Abstract

This application relates to the field of satellite signal processing technology and discloses a method and system for intelligent identification and adaptive acquisition of multi-constellation satellite signals. The method includes: extracting the frequency trajectory matrix by performing a short-time Fourier transform on the received signal and calculating the Doppler frequency shift rate and acceleration; combining modulation domain features with the input of a multi-task neural network to output constellation category and interference type identification results; when a low-Earth orbit satellite is identified, predicting the carrier frequency offset trajectory and uncertainty through a long short-time memory network; adjusting the frequency of the numerically controlled oscillator in advance based on the prediction results and dynamically setting the search window and integration time; implementing corresponding anti-interference strategies for different interference types; and periodically monitoring signal quality indicators, triggering strategy upgrades when the performance improvement rate falls below a threshold. This application improves the signal acquisition success rate and tracking stability in a multi-constellation mixed environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite signal processing, and particularly relates to a multi-constellation satellite signal intelligent identification and adaptive acquisition method and system. BACKGROUND

[0002] With the rapid development of global satellite navigation systems and low-orbit communication constellations, multiple satellite systems such as GPS, Beidou, Galileo and StarLink are running on the space orbit at the same time. These satellite signals overlap and interweave in the frequency spectrum, and the receiver needs to realize signal identification and acquisition in a complex multi-source signal mixed environment. The satellite signal receiver in the prior art usually adopts an identification method based on signal modulation characteristics, and distinguishes signals of different constellations by analyzing modulation domain characteristics such as power spectral density, chip rate and frame structure. After identification, signal acquisition is performed according to preset fixed parameters, and the acquisition parameters include frequency search range, integration time and the like. These parameters are set at the system initialization and remain unchanged. For the frequency deviation problem in the signal acquisition process, the prior art adopts a frequency lock loop for tracking compensation. When the carrier frequency deviates from the preset value, the loop adjusts the frequency of the local oscillator to follow the signal frequency change. This method belongs to a feedback control mechanism and performs post-correction after the frequency deviation occurs.

[0003] However, the prior art has the following deficiencies: First, the identification method based on modulation domain characteristics cannot effectively distinguish satellite signals of different orbital altitudes. Although the orbital parameters of StarLink low-orbit satellites and GPS medium-orbit satellites are significantly different, the modulation methods of some signals are similar, and only relying on power spectrum and code rate characteristics can easily cause misjudgment, especially under the condition of low signal-to-noise ratio, the modulation domain feature extraction accuracy decreases, and the identification accuracy is insufficient. Second, the fixed parameter acquisition strategy lacks adaptability and cannot adjust the processing parameters according to the actual dynamic characteristics of the signal. The low-orbit satellite has a high Doppler frequency shift rate of tens of thousands of hertz per second due to its low orbital altitude and high running speed, while the medium-orbit satellite has a frequency shift rate of only thousands of hertz per second. Using a unified frequency search window and integration time will result in low-orbit satellite acquisition efficiency or medium-orbit satellite processing gain deficiency. Third, the post-compensation frequency tracking method has inherent delay. The frequency lock loop needs to go through multiple control periods to eliminate the frequency deviation. In the high dynamic scene, the signal frequency changes rapidly, and the response lag of the loop results in frequent loss of lock, especially when the low-orbit satellite passes the zenith, the Doppler frequency shift acceleration is large, and the contradiction between the loop bandwidth and the tracking precision is difficult to balance.

[0004] Because dynamic characteristics such as Doppler frequency shift rate and acceleration are not extracted, the system cannot determine whether the signal belongs to a low-orbit satellite with high dynamics or a medium-orbit satellite with low dynamics, which limits the recognition accuracy and cannot perform targeted parameter configuration. Even if the signal category is identified, the fixed parameter strategy cannot respond to signal dynamics. When the low-orbit satellite signal changes rapidly due to rapid movement, the preset narrow frequency search window cannot cover the true frequency deviation range, and the preset long integration time will cause the correlation peak to spread and energy loss due to frequency walking. In addition, the post-compensation mechanism can only start adjusting after the frequency deviation has occurred and been detected. The adjustment process takes time, during which the signal frequency continues to change, and the compensation always lags behind the actual deviation, forming a situation that cannot be caught up with. This is particularly serious in the high-dynamic scenario of low-orbit satellites. Frequent loss of lock and recapture not only reduces the continuity of positioning services, but also increases power consumption and computational burden. The deeper problem is that existing technologies do not consider interference factors in a multi-constellation mixed environment. Signals in different frequency bands are affected differently by rain attenuation, adjacent frequency interference, and malicious interference. There is a lack of interference type identification and targeted anti-interference strategies, resulting in a significant decrease in signal acquisition success rate in complex electromagnetic environments. Even if successful acquisition is achieved, it is difficult to maintain stable tracking for a long time. SUMMARY

[0005] The present application provides a multi-constellation satellite signal intelligent identification and adaptive acquisition method and system, which extracts dynamic characteristics such as Doppler frequency shift rate and frequency shift acceleration, combines a multi-task neural network to realize joint identification of constellation category and interference type, and predicts future carrier frequency offset trajectories based on an LSTM time series prediction network. According to the prediction results, the frequency of the digital controlled oscillator is adjusted in advance to achieve predictive compensation. At the same time, the frequency search window and integration time are dynamically set according to the prediction uncertainty and frequency shift rate, and differential anti-interference strategies are executed for different interference types. The problems of low multi-constellation signal recognition accuracy, lack of adaptive capability of acquisition parameters, frequency tracking response lag, and insufficient anti-interference capability in existing technologies are solved, and the signal acquisition success rate and tracking stability in a multi-constellation mixed environment are improved.

[0006] In a first aspect, the present application provides a multi-constellation satellite signal intelligent identification and adaptive acquisition method, which comprises:

[0007] S1 step: performing short-time Fourier transform on the received signal, extracting a frequency-time trajectory matrix, calculating the Doppler frequency shift rate and frequency shift acceleration, and combining the modulation domain features to form a feature vector;

[0008] S2 step: inputting the feature vector into a multi-task neural network, outputting constellation category identification results and interference type classification results through shared feature extraction layers and double-branch task heads;

[0009] Step S3: When a low-Earth orbit satellite signal is identified, the carrier frequency offset trajectory within the future time window is predicted based on the historical frequency sequence using a long short-term memory network, and the prediction uncertainty is output simultaneously.

[0010] Step S4: Adjust the frequency of the numerically controlled oscillator in advance according to the carrier frequency offset trajectory, dynamically set the frequency search window range according to the prediction uncertainty, and adaptively adjust the integration time according to the frequency shift rate;

[0011] Step S5: Execute the corresponding anti-interference strategy according to the interference type classification result, extend the integration time and switch to high elevation angle satellite for rain attenuation interference, configure notch filter for adjacent channel interference, and perform constellation frequency band switching for malicious interference.

[0012] S6 Step: Periodically monitor carrier-to-noise ratio, bit error rate, and lockout duration. When the performance improvement rate falls below a set threshold, trigger a policy upgrade.

[0013] Secondly, this application provides a multi-constellation satellite signal intelligent identification and adaptive acquisition system, the multi-constellation satellite signal intelligent identification and adaptive acquisition system comprising:

[0014] The transformation module is used to perform short-time Fourier transform on the received signal, extract the frequency-time trajectory matrix, calculate the Doppler frequency shift rate and frequency shift acceleration, and combine the modulation domain features to form a feature vector;

[0015] The input module is used to input the feature vector into the multi-task neural network, and output constellation category recognition results and interference type classification results through a shared feature extraction layer and a dual-branch task head;

[0016] The prediction module is used to predict the carrier frequency offset trajectory within a future time window based on historical frequency sequences and a long short-term memory network when low-Earth orbit satellite signals are identified, while outputting the prediction uncertainty.

[0017] The adjustment module is used to adjust the frequency of the numerically controlled oscillator in advance according to the carrier frequency offset trajectory, dynamically set the frequency search window range according to the prediction uncertainty, and adaptively adjust the integration time according to the frequency shift rate.

[0018] The switching module is used to execute corresponding anti-interference strategies based on the interference type classification results, extend the integration time and switch to high elevation angle satellites for rain attenuation interference, configure notch filters for adjacent channel interference, and perform constellation frequency band switching for malicious interference.

[0019] The upgrade module is used to periodically monitor the carrier-to-noise ratio, bit error rate, and lockout duration. When the performance improvement rate is lower than a set threshold, a policy upgrade is triggered.

[0020] Thirdly, a multi-constellation satellite signal intelligent identification and adaptive acquisition device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the multi-constellation satellite signal intelligent identification and adaptive acquisition device to execute the above-described multi-constellation satellite signal intelligent identification and adaptive acquisition method.

[0021] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, cause the computer to perform the above-described method for intelligent identification and adaptive acquisition of multi-constellation satellite signals.

[0022] The technical solution provided in this application extracts the frequency-time trajectory matrix by performing a short-time Fourier transform on the received signal and calculates the Doppler frequency shift rate and frequency shift acceleration. Combined with modulation domain features, a feature vector is constructed, overcoming the limitations of existing technologies that rely solely on modulation domain features for identification. The introduction of dynamic features enables the system to distinguish the motion characteristics of satellites at different orbital altitudes. Low-Earth orbit satellites exhibit high frequency shift rates and high accelerations due to their small orbital radius and high angular velocity, while medium- and high-Earth orbit satellites exhibit low frequency shift rates and low accelerations due to their large orbital radius and low angular velocity. This feature extraction method based on physical motion laws maintains stable discrimination capability even under low signal-to-noise ratio conditions. The feature vector is input into a multi-task neural network, which outputs constellation category identification results and interference type classification results through a shared feature extraction layer and a dual-branch task head. The multi-task learning architecture simultaneously completes constellation identification and environmental perception in a single forward propagation. Shared low-level feature extraction reduces redundant computation, and the dual-branch structure allows the two tasks to be independent yet share knowledge. The constellation identification result guides whether to activate the prediction module, and the interference identification result determines the selection of anti-interference strategies, achieving an organic integration of signal identification and environmental perception. When low-Earth orbit (LEO) satellite signals are identified, the carrier frequency offset trajectory within the future time window is predicted based on historical frequency sequences using a long short-term memory (LSTM) network, and the prediction uncertainty is output. The gating mechanism of the LSTM network enables it to learn the time dependence of the frequency sequence. The forget gate selectively retains long-term trend information, the input gate absorbs the change characteristics of the current moment, and the output gate integrates historical and current information to generate the prediction result. The attention mechanism further enhances the weight of key moments, and the quantification of prediction uncertainty provides a reliability assessment basis for subsequent parameter adjustments. Compared with traditional polynomial fitting or Kalman filtering, this prediction method based on time series modeling can more accurately capture the nonlinear frequency changes when LEO satellites pass overhead. The numerically controlled oscillator frequency is adjusted in advance based on the carrier frequency offset trajectory, transforming traditional post-compensation into predictive compensation. This eliminates the response delay of the frequency tracking loop, ensuring the local carrier is in place before the signal frequency shifts, thus avoiding frequent loss of lock-up caused by catch-up tracking. The frequency search window range is dynamically set based on prediction uncertainty: narrowing the search range to reduce computation when prediction is reliable, and expanding it to ensure coverage of the true frequency offset when prediction is uncertain. The integration time is adaptively adjusted based on the frequency shift rate: shortening the integration time in high dynamics to avoid correlation peak broadening, and extending it in low dynamics to increase processing gain. This multi-parameter adaptive mechanism overcomes the rigidity of fixed-parameter strategies. Corresponding anti-interference strategies are executed based on interference type classification results. For rain attenuation interference, the integration time is extended to compensate for power loss and a switch to a high-elevation satellite is used to shorten the propagation path. For adjacent-channel interference, notch filters are configured to precisely suppress interfering frequencies. For malicious interference, constellation band switching is performed to avoid broadband interference. The hierarchical decision tree architecture enables the system to take differentiated measures against different interference sources, significantly improving survivability in complex electromagnetic environments compared to single anti-interference methods.The system periodically monitors the carrier-to-noise ratio, bit error rate, and lock-in duration. When the performance improvement rate falls below a set threshold, a policy upgrade is triggered, forming a closed-loop adaptive mechanism from parameter adjustment to effect evaluation to policy optimization. The system can not only make decisions based on the current state, but also continuously optimize based on the effect of the decisions. This closed-loop feedback enables the system to have the ability to learn and evolve on its own.

[0023] In specific application scenarios of intelligent identification and adaptive acquisition of multi-constellation satellite signals, the algorithm features of this application make key contributions to the solution. The dynamic feature parameters extracted by combining short-time Fourier transform with first-order and second-order difference operations have physical meanings that directly correspond to the satellite orbital dynamics characteristics, giving the feature extraction process clear theoretical support rather than being purely data-driven. The multi-task neural network achieves dual recognition function with only a limited increase in the number of parameters and computation through the architecture design of shared convolutional layers and dual-branch Softmax layers. During network training, the loss functions of the two tasks are weighted and combined to form a joint optimization objective, so that the constellation recognition task and the interference recognition task promote each other rather than interfere with each other. The forget gate, input gate, and output gate triple gating mechanism of the LSTM network enable it to adaptively adjust the degree of dependence on information from different historical moments throughout the entire satellite overpass process. In the early stage of the overpass, when the horizontal velocity component is large and the frequency shift rate changes drastically, the input gate has a high weight to absorb the current changes. In the middle stage of the overpass, when the vertical velocity component is large and the frequency shift rate changes are stable, the forget gate has a high weight to retain the long-term trend. The attention mechanism automatically discovers the most critical time step for prediction through a learnable weight matrix, avoiding the subjectivity of manually designing the weighting function. Prediction uncertainty is estimated using a Monte Carlo sampling method with multiple forward propagations, reflecting the network's confidence in its own predictions. This confidence assessment enables subsequent parameter adjustments to have risk awareness capabilities. A conservative strategy is adopted to expand the search range when uncertainty is high, while an aggressive strategy is adopted to narrow the search range when uncertainty is low, achieving a dynamic balance between risk and efficiency. The transfer function coefficients of the notch filter are calculated in real time based on the normalized angular frequency and notch depth control parameters, generating customized filter response curves for different interference frequencies. Compared to a scheme with a pre-set fixed filter bank, this approach offers greater flexibility and higher spectral efficiency. The multi-stage cascaded notch filter accumulates the suppression depth by connecting multiple single-stage notch units in series, with each additional stage increasing the notch depth by approximately twenty decibels, enabling the system to cope with strong interference scenarios. The incremental learning mechanism selectively stores high-accuracy samples and freezes the weights of the shared layer, only fine-tuning the task branch layer. While retaining the original knowledge, it absorbs the feature distribution of the new scene, avoiding the catastrophic forgetting problem. This allows the model to continuously improve its performance as the runtime increases. This online learning capability gives the system the potential for long-term evolution to adapt to environmental changes. These features of the algorithm together constitute the technical advantages of this application in the field of multi-constellation signal processing. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of an embodiment of the intelligent identification and adaptive acquisition method for multi-constellation satellite signals in this application.

[0026] Figure 2 This is a schematic diagram illustrating the frequency search window range under different reliability levels in the embodiments of this application;

[0027] Figure 3 This is a schematic diagram of one embodiment of the intelligent identification and adaptive acquisition system for multi-constellation satellite signals in this application.

[0028] Figure 4 This is a schematic block diagram of the structure of the intelligent identification and adaptive acquisition device for multi-constellation satellite signals in an embodiment of the present invention. Detailed Implementation

[0029] This application provides a method and system for intelligent identification and adaptive acquisition of multi-constellation satellite signals. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0030] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent identification and adaptive acquisition method for multi-constellation satellite signals in this application includes:

[0031] Step S1: Perform a short-time Fourier transform on the received signal, extract the frequency-time trajectory matrix, calculate the Doppler frequency shift rate and frequency shift acceleration, and combine the modulation domain features to form a feature vector;

[0032] Step S2: Input the feature vector into the multi-task neural network, and output the constellation category recognition result and interference type classification result through the shared feature extraction layer and the dual-branch task head;

[0033] Step S3: When a low-Earth orbit satellite signal is identified, the carrier frequency offset trajectory within the future time window is predicted based on the historical frequency sequence using a long short-term memory network, and the prediction uncertainty is output simultaneously.

[0034] Step S4: Adjust the CNC oscillator frequency in advance according to the carrier frequency offset trajectory, dynamically set the frequency search window range according to the prediction uncertainty, and adaptively adjust the integration time according to the frequency shift rate.

[0035] S5 step: Execute the corresponding anti-interference strategy according to the interference type classification result. For rain attenuation interference, extend the integration time and switch to high elevation angle satellites. For adjacent channel interference, configure notch filters. For malicious interference, perform constellation frequency band switching.

[0036] S6 Step: Periodically monitor carrier-to-noise ratio, bit error rate, and lockout duration. When the performance improvement rate falls below a set threshold, trigger a policy upgrade.

[0037] It is understood that the executing entity of this application can be a multi-constellation satellite signal intelligent identification and adaptive acquisition system, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0038] Specifically, the receiver front-end performs analog-to-digital conversion and digital down-conversion on the original RF signal, shifting the high-frequency signal to zero intermediate frequency (IF) to generate a baseband complex signal with in-phase and quadrature components. A short-time Fourier transform is applied to the baseband complex signal, using a Hamming window as the time window function. The window length is set to 2048 sampling points, and the overlap rate between adjacent windows is set to 75%, meaning the window slides by 512 sampling points each time. The transform result forms a two-dimensional time-frequency matrix, with rows corresponding to the frequency axis and columns corresponding to the time axis. The maximum spectral energy point is searched within each time frame of this matrix, and the corresponding frequency value is recorded. The peak frequencies of consecutive time frames are arranged in chronological order to form a frequency sequence. A first-order difference operation is performed on this frequency sequence to calculate the frequency change between adjacent moments. This change is divided by the time interval to obtain the Doppler frequency shift rate, where the time interval is equal to the time length corresponding to the window sliding step. A second-order difference operation is then performed on the frequency shift rate sequence to calculate the rate change and divide it by the time interval to obtain the frequency shift acceleration. The baseband complex signal is subjected to a fourth-order operation, followed by a Fast Fourier Transform (FFT). The peak intensity of the spectrum at four times the carrier frequency is detected on the frequency axis, and this intensity value is used as a modulation domain characteristic parameter. The Doppler shift rate reflects the rate of frequency change caused by the satellite's radial velocity relative to the receiver. Low-Earth orbit (LEO) satellites, due to their low altitude and high speed, can have frequency shift rates ranging from tens of thousands to hundreds of thousands of hertz per second, while medium- and high-Earth orbit (MEO) satellites have frequency shift rates of only a few thousand hertz per second. Frequency shift acceleration reflects the rate of velocity change; LEO satellites experience drastic acceleration changes when passing overhead. The Doppler shift rate, frequency shift acceleration, and modulation domain characteristic parameters are arranged sequentially to form a feature vector. This feature vector contains the signal's dynamic and modulation characteristics. The dynamic characteristics are used to distinguish satellites at different orbital altitudes, while the modulation characteristics are used to distinguish signals with different modulation schemes.

[0039] The feature vector is input into the first one-dimensional convolutional layer of the multi-task neural network. This layer contains 64 convolutional kernels, each with a length of 5. A sliding convolution operation is performed on the input vector, multiplying the kernels element-wise with local regions of the input data, summing the results, and adding a bias term. The result is then processed by the ReLU activation function, which sets negative values ​​to zero and retains positive values. The first convolutional layer outputs 64 feature maps. The second convolutional layer has 128 kernels with a length of 3, continuing the convolution operation on the 64 feature maps from the first layer to output 128 feature maps. The third convolutional layer has 256 kernels, resulting in 256 feature maps after the three convolutional layers. Each feature map corresponds to a different feature representation. The role of the convolutional layers is to extract local patterns and feature combinations from the input data. A global average pooling layer calculates the average value for each of the 256 feature maps, summing all elements of each feature map and dividing by the total number of elements to obtain 256 scalar values. These 256 scalar values ​​form a shared feature vector. The pooling operation reduces the data dimensionality while preserving the main feature information. The shared feature vectors are input to two task branches simultaneously. The constellation recognition task branch contains two fully connected layers. The first fully connected layer has 128 neurons, each performing a fully connected operation with the input 256-dimensional vector. Each input component is multiplied by its corresponding weight, summed, and then a bias is added. The result is activated by ReLU. The second fully connected layer has 3 neurons corresponding to the Starlink, GPS, and BeiDou categories. After outputting three values, they are normalized by a Softmax layer. Softmax converts the original values ​​into a probability distribution, with each category corresponding to a probability value, and the sum of all probability values ​​equals 1. The interference recognition task branch has a similar structure. The last layer has 4 neurons corresponding to the rain attenuation, adjacent frequency, malicious, and no interference categories, and also outputs a four-dimensional probability vector through Softmax. In the constellation category probability vector, the values ​​of the three components are compared, and the category corresponding to the component with the largest value is selected as the constellation category recognition result. In the interference type probability vector, the category corresponding to the highest probability is selected as the interference type classification result. The multi-task neural network reduces the computational load by sharing the underlying feature extraction, and the dual-branch structure enables parallel processing of constellation recognition and interference recognition.

[0040] The system determines the category label of the constellation identification result. When the label is Starlink or other low-Earth orbit satellite types, it extracts the most recent frequency data from the frequency sequence as a historical frequency sequence. The extracted data corresponds to the past 200 milliseconds and contains 24 frequency sampling points. The mean of the historical frequency sequence is calculated by summing all frequency values ​​and dividing by the sequence length. To calculate the standard deviation, the system first calculates the squared difference between each frequency value and the mean, sums all the squared differences, divides by the sequence length, and then takes the square root of the result. Each element of the historical frequency sequence is normalized by subtracting the mean from the original frequency value and then dividing by the standard deviation. The normalized values ​​are distributed within a range of zero mean and unit variance. Normalization eliminates the influence of different units, facilitating neural network training. A normalized frequency sequence is input into an LSTM network. Each LSTM unit contains three gate structures: the forget gate, which determines the proportion of information discarded from the cell state; the forget gate concatenates the previous hidden state and the current input into a vector, multiplies it by the weight matrix, adds a bias, and then passes it through a Sigmoid activation function. The Sigmoid function outputs a value between 0 and 1, representing the retention proportion. The input gate determines the cell state update proportion, also calculated using a Sigmoid function, and generates candidate values ​​using a Tanh activation function, which outputs a value between -1 and +1. Updating the cell state involves multiplying the old cell state by the forget gate output, adding the product of the input gate output and the candidate value. The output gate determines the hidden state output, calculating the output proportion using a Sigmoid function, processing the cell state through a Tanh function, and multiplying it by the output gate output to obtain the new hidden state. The LSTM network consists of two layers: the first layer has 128 units, and the second layer has 64 units. The input at each time step is processed sequentially, and the hidden states from all time steps are finally input into the attention mechanism layer. The attention mechanism calculates the importance weights at each time step. First, each hidden state is transformed using a learnable weight matrix, then normalized using the Softmax function to obtain the attention weights. The weighted context vector is obtained by multiplying all hidden states by their corresponding weights and summing the results. This context vector is input into a fully connected layer, which consists of three layers with 64, 32, and 12 neurons respectively. The last layer outputs 12 predicted values, corresponding to frequency predictions approximately every 8 milliseconds within the next 100 milliseconds. The predicted values ​​are then denormalized by multiplying them by the previously calculated standard deviation and adding the mean to restore them to the original frequency units. The denormalized prediction sequence constitutes the carrier frequency offset trajectory. Prediction uncertainty is calculated through multiple forward propagations. During network inference, forward propagation is performed multiple times, obtaining a set of predicted values ​​each time. The standard deviation of these multiple sets of predictions is calculated at each time point. The standard deviation reflects the degree of prediction fluctuation, i.e., uncertainty; a higher uncertainty value indicates lower prediction reliability.

[0041] The predicted frequency value for a future moment is extracted from the carrier frequency offset trajectory. This moment is typically chosen as the prediction point 50 milliseconds in the future. The frequency setpoint of the numerically controlled oscillator (CNC) is obtained by subtracting this predicted frequency value from the intermediate frequency (IF). The CNC generates a local carrier signal of a specific frequency through a phase accumulator. The frequency setpoint is divided by the system clock frequency to obtain the normalized frequency, which is then multiplied by the value corresponding to the bit width of the phase accumulator to obtain the phase increment. The bit width of the phase accumulator determines the frequency resolution; a larger bit width results in higher frequency control accuracy. The phase increment value is loaded into the phase accumulator register of the CNC oscillator. The accumulator adds the phase increment to the current phase in each clock cycle. The accumulated phase value is mapped to a sine-cosine lookup table to output in-phase and quadrature components, thus realizing the frequency adjustment of the local carrier. The prediction reliability level is determined based on the prediction uncertainty. Two thresholds are set to divide the uncertainty into three intervals. When the uncertainty is below the first threshold, the prediction reliability is high, and a narrow first frequency search window range is set. When the uncertainty is between the first and second thresholds, the prediction reliability is medium, and a medium-width second frequency search window range is set. When the uncertainty is above the second threshold, the prediction reliability is low, and a wide third frequency search window range is set. The search window range defines the upper and lower limits of frequency search during the acquisition process. The wider the range, the more frequencies need to be searched, but it can accommodate larger frequency deviations. The signal dynamic level is determined based on the frequency shift rate. Two rate thresholds are set to divide the dynamic level into three levels. When the absolute value of the frequency shift rate is above the third threshold, the signal dynamics are fast, and a short first integration time is set. A short integration time can avoid correlation peak broadening and energy loss caused by rapid frequency changes. When the absolute value of the frequency shift rate is between the third and fourth thresholds, the signal dynamics are medium, and a medium second integration time is set. When the absolute value of the frequency shift rate is below the fourth threshold, the signal dynamics are slow, and a longer third integration time is set. A longer integration time can accumulate more energy, improve processing gain, and improve acquisition performance under low signal-to-noise ratio. The adjusted CNC oscillator frequency setting, frequency search window range, and integration time are combined to form a set of capture parameters, which are then passed to the capture execution module to control the subsequent signal capture process.

[0042] The system determines whether the interference type is rain attenuation interference, adjacent channel interference, or malicious interference. If the result is rain attenuation interference, the system reads the signal-to-noise ratio (SNR) values ​​from the current moment and multiple historical moments, calculates the difference sequence between the current and historical SNR, and performs linear fitting on the difference sequence to obtain the slope of the SNR change over time, i.e., the SNR decrease rate. Rain attenuation is characterized by a continuous and slow decrease in SNR. When the SNR decrease rate exceeds the rain attenuation threshold, rain attenuation interference is confirmed. The integration time is extended by a set percentage based on the value set in step S4. Extending the integration time increases the signal accumulation time, which can partially compensate for the signal power loss caused by rain attenuation. The system queries satellite ephemeris data to obtain the orbital parameters of the currently tracked satellite. Based on the receiver position and the satellite position, the system calculates the elevation angle of the currently tracked satellite, which is the angle between the satellite's direction and the ground plane. When the elevation angle is lower than the elevation angle threshold, it indicates a longer signal propagation path and more severe rain attenuation. The system searches the visible satellite list for alternative satellites with elevation angles higher than this threshold. Satellites with higher elevation angles have shorter propagation paths and less rain attenuation loss. A satellite switching command is sent to the RF front-end module to complete the satellite switching. When the judgment result is adjacent-channel interference, the power spectral density curve is extracted from the time-frequency two-dimensional matrix generated in step S1. This curve describes the distribution of signal power on the frequency axis. Peak detection is performed on the power spectral density curve. The interference detection threshold is set as the average power plus a certain margin. All frequency points exceeding this threshold are recorded to form an interference frequency set. These frequency points correspond to the center frequency of the adjacent-channel interference source. A second-order IIR notch filter is designed for each interference frequency point in the interference frequency set. The normalized angular frequency is calculated as the ratio of the interference frequency point to the sampling frequency multiplied by 2 times pi. The notch depth control parameter determines the depth and bandwidth of the notch. The closer the parameter value is to 1, the deeper the notch but the narrower the bandwidth. The transfer function of the filter is generated based on the normalized angular frequency and the notch depth control parameter. The numerator and denominator coefficients of the transfer function describe the frequency response characteristics of the filter. These coefficients are loaded into the register of the configurable filter module. The filter module performs real-time filtering processing on the baseband signal to suppress adjacent-channel interference. When the judgment result is malicious interference, the difference between the current signal-to-noise ratio (SNR) and the previous SNR is calculated to obtain the instantaneous drop amplitude. The characteristic of malicious interference is a sudden and significant drop in SNR. When the instantaneous drop amplitude exceeds the malicious interference judgment threshold, malicious interference is confirmed to exist. The frequency band configuration tables of different constellations are queried to obtain the operating frequency band information of each constellation system. A candidate constellation with a frequency band difference exceeding the frequency band switching threshold is selected. A large frequency band difference means that the power spectral density of the interference is significantly reduced in the new frequency band. A local oscillator frequency switching command is sent to the RF front-end module to adjust the local oscillator frequency to the frequency band corresponding to the new constellation. At the same time, a signal mode switching command is sent to adjust the modulation and demodulation mode and frame structure parameters of the baseband processing module to match the signal format of the new constellation.

[0043] Signal quality monitoring is performed periodically, with three key metrics calculated for each monitoring cycle. First, the ratio of the correlated peak power to the noise floor power is calculated. The correlated peak power is obtained by squared the amplitude of the peak point in the correlation operation, while the noise floor power is estimated by the power variance of non-peak points surrounding the correlated peak. This ratio reflects the relative strength of the signal to the noise. Second, the carrier-to-noise ratio (CNR) is calculated based on this ratio and the equivalent noise bandwidth. The CNR is the logarithm of the ratio plus the logarithm of the bandwidth, and its unit is dB / Hertz, describing the ratio of signal power to noise power per unit bandwidth. Third, the number of errors in pilot symbols during the hard decision process is counted. Hard decision maps the received symbol to the nearest standard symbol point after comparing it with the decision threshold. The hard decision result is compared symbol by symbol with the known pilot symbol sequence, and the number of inconsistencies is counted. The number of erroneous symbols divided by the total number of symbols yields the bit error rate (BER), which reflects the correctness of signal demodulation. Fourth, the lock-on time is recorded from successful signal acquisition to tracking loss, reflecting the stability of the receiver in maintaining the signal lock state. The measured carrier-to-noise ratio (CNR), bit error rate (BER), and lock-and-hold time are compared with the baseline CNR, BER, and lock-and-hold time recorded before policy execution. The change in each indicator is calculated by subtracting the baseline value from the current value, and dividing the change by the baseline value yields the improvement rate for that indicator. The combined improvement rates of the three indicators are used to calculate the performance improvement rate, with a positive rate indicating performance improvement and a negative rate indicating performance degradation. The system checks if the performance improvement rate is below a set threshold. If the improvement rate is below the threshold and the duration of this state exceeds the judgment time, a policy upgrade mechanism is triggered. The duration judgment avoids false triggering caused by instantaneous fluctuations. Based on the interference type classification results, the corresponding upgrade strategy is selected. For rain-attenuated interference, the integration time is further extended, and a command to reduce the coding rate is sent forward to the error correction coding module. The reduced code rate increases redundancy information and improves error correction capability. For adjacent-channel interference, the number of cascaded notch filters is increased. Multi-stage cascaded filters connect more notch units in series with the original single-stage notch filter, and each notch filter independently suppresses the cumulative interference, enhancing the notch depth. To address malicious interference, the system switches to a secondary alternative constellation. If interference persists even after switching to the first alternative constellation, it switches to the second, and so on, trying each level until a constellation system with lower interference is found. Sample data with recognition accuracy exceeding a threshold across multiple consecutive evaluation periods are stored in a local sample buffer. Recognition accuracy is calculated by comparing the recognition results with the actual signal type. The buffer records feature vectors, true labels, and predicted labels. When the local sample buffer reaches its capacity limit, a model fine-tuning process is triggered. The buffer samples are mixed with the original training set in a set ratio to form a new training dataset. This ratio control ensures a certain proportion of new samples while retaining existing training samples to avoid catastrophic forgetting.The weights of the task branch layers of the multi-task neural network are fine-tuned. The weights of the shared feature extraction layer are frozen and not updated. Only the weights of the fully connected layers of the constellation recognition branch and the interference recognition branch are updated. The gradient of the loss function with respect to the weights is calculated by the backpropagation algorithm and the weight values ​​are adjusted according to the gradient. After the fine-tuning is completed, the updated network parameters are obtained and the original parameters are replaced.

[0044] In one specific embodiment, step S1 includes:

[0045] The received signal is digitally down-converted to obtain a baseband complex signal with in-phase and quadrature components;

[0046] A short-time Fourier transform is performed on the baseband complex signal to obtain a two-dimensional time-frequency matrix;

[0047] Search for the peak points of spectral energy in each time slice of the time-frequency two-dimensional matrix, record the peak frequencies, and obtain the frequency sequence;

[0048] Perform a first-order difference operation on the frequency sequence and divide it by the time interval to obtain the Doppler frequency shift rate. Perform a second-order difference operation on the Doppler frequency shift rate and divide it by the time interval to obtain the frequency shift acceleration.

[0049] After performing a fourth power operation on the baseband complex signal, a fast Fourier transform is performed, and the peak intensity is detected at four times the carrier frequency to obtain the modulation domain characteristics.

[0050] The Doppler frequency shift rate, frequency shift acceleration, and modulation domain characteristics are used to obtain the feature vector.

[0051] Specifically, the receiver front-end samples the radio frequency signal through an analog-to-digital converter to obtain a digital signal. Digital down-conversion processing multiplies this digital signal with locally generated sine and cosine carriers to obtain the in-phase component I and the quadrature component Q, respectively. The in-phase component corresponds to the signal multiplied by the in-phase carrier and then low-pass filtered, and the quadrature component corresponds to the signal multiplied by the quadrature carrier and then low-pass filtered. These two components together constitute the baseband complex signal, with the real part being the in-phase component and the imaginary part being the quadrature component. Short-time Fourier transform (SFT) segments the baseband complex signal. A fixed-length time window is selected to capture a segment of data starting from the signal's beginning. A fast Fourier transform is performed on this segment to obtain the spectral distribution within the time window. The window slides forward according to a set step size to capture the next segment of data and perform the Fourier transform again. This process is repeated until the entire signal is processed. The spectral results of all time windows are arranged in chronological order to form a two-dimensional time-frequency matrix, where the row index corresponds to the frequency point and the column index corresponds to the time window. In each time slice (column) of the two-dimensional time-frequency matrix, the frequency point with the largest amplitude is searched. All rows in that column are traversed to find the row index with the largest amplitude value. The frequency value corresponding to this row index is the peak frequency of that time slice. The peak frequencies of all time slices are recorded in chronological order to form a frequency sequence, which reflects the trajectory of the frequency where the main energy of the signal is located over time. The first-order difference operation of the frequency sequence takes the difference between two adjacent frequency values. The second frequency value is subtracted from the first frequency value to obtain the frequency change. This change is divided by the time interval between the two sampling points to obtain the frequency change rate, i.e., the Doppler frequency shift rate, within that time interval. The time interval is equal to the time length corresponding to the window sliding step. Point-by-point difference is performed on the entire frequency sequence to obtain the frequency shift rate sequence. The second-order difference operation of the Doppler frequency shift rate sequence also takes the difference between two adjacent rate values. Subtracting the first rate value from the second rate value yields the rate change. Dividing this change by the time interval gives the rate of change, i.e., the frequency shift acceleration. Frequency shift acceleration reflects how fast the Doppler frequency shift rate changes. Low-Earth orbit satellites have high Doppler frequency shift rates and large accelerations due to their small orbital radius and large angular velocity, while medium- and high-Earth orbit satellites have low frequency shift rates and small accelerations due to their large orbital radius and small angular velocity. The fourth-order operation of the baseband complex signal multiplies each sample point of the complex signal by four times. The fourth-order operation shifts the signal spectrum to four times the carrier frequency in the frequency domain. The fourth-order spectrum of the BPSK modulated signal produces a significant peak at the fourth-order frequency point. The peak position of the fourth-order spectrum of the QPSK modulated signal is different. Performing a Fast Fourier Transform on the fourth-order signal yields its spectral distribution. Finding the frequency point corresponding to four times the carrier frequency on the frequency axis and reading the amplitude value of this frequency point is taken as the peak intensity. This peak intensity reflects the modulation characteristics of the signal.The Doppler frequency shift rate, frequency shift acceleration, and modulation domain characteristic peak intensity values ​​are arranged in order to form a one-dimensional array to form a feature vector. This feature vector contains both the time-varying dynamic characteristics and frequency domain modulation characteristics of the signal. The frequency shift rate and acceleration of the dynamic characteristics are used to distinguish satellite constellations at different orbital altitudes, and the modulation domain characteristics are used to distinguish signal systems with different modulation methods.

[0052] In one specific embodiment, step S2 includes:

[0053] The feature vector is input into a one-dimensional convolutional neural network layer of a multi-task neural network for feature extraction. Through multi-layer convolution operations and activation function processing, a shared feature vector is obtained.

[0054] The shared feature vector is input into a global average pooling layer for dimensionality reduction, resulting in a fixed-dimensional feature representation.

[0055] The fixed-dimensional feature representations are simultaneously input into the constellation recognition task branch and the interference recognition task branch. The constellation recognition task branch contains a constellation fully connected layer and a constellation softmax layer, and outputs a constellation category probability vector. The interference recognition task branch contains an interference fully connected layer and an interference softmax layer, and outputs an interference type probability vector.

[0056] By comparing the probability values ​​of each component in the constellation category probability vector, the category label corresponding to the highest probability value is selected to obtain the constellation category recognition result. Similarly, by comparing the probability values ​​of each component in the interference type probability vector, the category label corresponding to the highest probability value is selected to obtain the interference type classification result.

[0057] Specifically, the feature vector is input into the first one-dimensional convolutional layer of the multi-task neural network. This convolutional layer contains multiple convolutional kernels, each a fixed-length array of weights. The kernels slide across the feature vector, multiplying the kernel's weights by the corresponding elements of the feature vector at each slide, summing the results, and adding a bias term to obtain an output value. After the kernels have slid to all possible positions, an output feature map is formed. The output feature maps of all convolutional kernels are then combined to form the output of this layer. The output is processed by the ReLU activation function, which sets negative values ​​to zero while leaving positive values ​​unchanged. This activation function introduces non-linearity, enabling the network to learn complex mapping relationships. The second convolutional kernel continues the convolution operation on the feature map output from the first layer, and the third convolutional kernel continues the convolution on the output of the second layer. After multiple convolutions, a higher-level feature representation is obtained. The features extracted by each convolutional layer become increasingly abstract, gradually combining local patterns into global features. The global average pooling layer calculates the average value for each feature map output from the convolutional layer. It sums all elements in the feature map and divides by the total number of elements, producing an average value for each feature map. The average values ​​of all feature maps form a fixed-dimensional shared feature vector. Pooling reduces data dimensionality while preserving the overall information of each feature map. This shared feature vector is simultaneously input to both the constellation recognition task branch and the interference recognition task branch. These two branches process data independently but share the results of the underlying feature extraction. The constellation recognition task branch contains fully connected layers. Each neuron in the fully connected layer is connected to all elements of the input vector, with each connection corresponding to a weight parameter. The output of a neuron equals the sum of the products of each element of the input vector and its corresponding weight, plus the bias. The output of the first fully connected layer is activated by ReLU and then input to the second fully connected layer. The number of neurons in the second fully connected layer equals the number of constellation categories, with each neuron corresponding to one constellation category. The output consists of three values ​​representing the scores for Starlink, GPS, and BeiDou, respectively. The Softmax layer converts the three scores into a probability distribution. During calculation, an exponential function is applied to each score, and the sum of the three exponential values ​​serves as a normalization factor. Each exponential value is divided by the normalization factor to obtain the probability of the corresponding category, and the sum of the three probability values ​​equals 1. The structure of the interference recognition task branch is similar to that of the constellation recognition branch. The last fully connected layer has four neurons corresponding to the four categories of rain attenuation, adjacent frequency, malicious, and no interference. After passing through the Softmax layer, it outputs a four-dimensional probability vector. In the constellation category probability vector, the three probability values ​​are compared, and the probability value with the largest value is found. The category index corresponding to this probability value is the constellation category recognition result. Similarly, in the interference type probability vector, the category index corresponding to the largest probability value is found as the interference type classification result.

[0058] In one specific embodiment, step S3 includes:

[0059] Determine whether the constellation category identification result is a low-Earth orbit satellite signal. If the result is yes, extract the historical frequency sequence from the frequency sequence.

[0060] The historical frequency sequence is normalized by calculating the mean and standard deviation. The mean is subtracted from each frequency value in the historical frequency sequence and then divided by the standard deviation to obtain the normalized frequency sequence.

[0061] The normalized frequency sequence is input into the Long Short-Term Memory network, and the cell state and hidden state are updated through gating operations of forget gate, input gate and output gate. After being weighted by the attention mechanism layer, it is input into the fully connected layer to obtain the normalized prediction sequence.

[0062] The normalized prediction sequence is denormalized, and each predicted value in the normalized prediction sequence is multiplied by the standard deviation and then the mean is added to obtain the carrier frequency offset trajectory. The standard deviation of the carrier frequency offset trajectory is calculated through multiple forward propagations to obtain the prediction uncertainty.

[0063] Specifically, the system determines whether the constellation category identification result belongs to low-Earth orbit (LEO) satellites. LEO satellites include Starlink and other satellite systems with orbital altitudes below 2000 km. When the category label is LEO satellite, the system extracts recent frequency data from the frequency sequence. The extraction length corresponds to a set historical time window containing several frequency sampling points, which constitute the historical frequency sequence. The mean of the historical frequency sequence is calculated by summing all frequency values ​​and dividing by the sequence length. To calculate the standard deviation, the mean is subtracted from each frequency value to obtain the deviation. The squares of all deviations are summed and divided by the sequence length to obtain the variance. The square root of the variance is then taken to obtain the standard deviation. Normalization is performed on each element of the historical frequency sequence. The mean is subtracted from each element, and the result is divided by the standard deviation to obtain the normalized value. The normalized value has a mean of zero and a standard deviation of 1. Normalization eliminates the differences in the dimensions and numerical ranges of the original data, making the neural network training more stable and converging faster. A normalized frequency sequence is input into the LSTM network. Each LSTM unit contains two internal states: a cell state and a hidden state. The cell state is used for long-term memory, while the hidden state is used for short-term output. The forget gate concatenates the hidden state from the previous time step with the current input into a vector. This vector is multiplied by the forget gate's weight matrix, a bias is added, and then it passes through the Sigmoid activation function. The Sigmoid function maps the input to a range of 0 to 1. The output value, representing the forgetting ratio, is multiplied element-wise by the cell state from the previous time step, retaining some old information while forgetting others. Similarly, the input gate multiplies the concatenated vector with its weight matrix, adds a bias, and passes it through the Sigmoid function to obtain the update ratio. Simultaneously, the concatenated vector is multiplied by the candidate value weight matrix, a bias is added, and then it passes through the Tanh activation function to obtain candidate values. The Tanh function maps the input to a range of -1 to +1. The update ratio is multiplied element-wise by the candidate values ​​to obtain the new information. The cell state update adds the old information processed by the forget gate to the new information generated by the input gate. The output gate multiplies the concatenated vector with the output gate weight matrix, adds a bias, and then passes it through the Sigmoid function to obtain the output ratio. The cell state is processed by the Tanh function and multiplied element-wise with the output ratio to obtain the new hidden state. The LSTM network processes the normalized frequency sequence step by step, updating the cell state and hidden state at each time step. The hidden states of all time steps are input into the attention mechanism layer. The attention mechanism calculates a weight for each hidden state, reflecting the importance of that time step to the prediction result. The weight calculation transforms the hidden state through a learnable weight matrix, and the transformation result is normalized by Softmax to obtain the attention weights. All hidden states are multiplied element-wise with their corresponding weights and summed to obtain the weighted context vector. The context vector is input into the fully connected layer for dimensionality transformation and nonlinear mapping. The last fully connected layer outputs the number of predicted time points, which correspond to the normalized frequency prediction values ​​at each time point within the future time window.Denormalization multiplies each predicted value by the previously calculated standard deviation and adds the mean to restore it to the original frequency range and unit. The denormalized prediction sequence constitutes the carrier frequency offset trajectory. Prediction uncertainty is calculated through multiple forward propagations. With the network's dropout mechanism enabled, some neurons are randomly discarded during each forward propagation. Multiple forward propagations yield multiple sets of prediction results. The standard deviation is calculated for the multiple sets of prediction values ​​at each prediction time point. The standard deviation quantifies the dispersion of the prediction, i.e., the uncertainty.

[0064] In one specific embodiment, step S4 includes:

[0065] The predicted frequency value for future moments is extracted based on the carrier frequency offset trajectory. The predicted frequency value is subtracted from the intermediate frequency to obtain the frequency setting value of the numerically controlled oscillator. The frequency setting value of the numerically controlled oscillator is divided by the system clock frequency and then multiplied by the value corresponding to the bit width of the phase accumulator to obtain the phase increment. The phase accumulator of the numerically controlled oscillator is adjusted according to the phase increment.

[0066] The prediction reliability level is determined based on the prediction uncertainty. When the prediction uncertainty is lower than the first threshold, a first frequency search window range is set. When the prediction uncertainty is between the first and second thresholds, a second frequency search window range is set. When the prediction uncertainty is higher than the second threshold, a third frequency search window range is set. The third frequency search window range is larger than the second frequency search window range, and the second frequency search window range is larger than the first frequency search window range.

[0067] The signal dynamic level is determined based on the frequency shift rate. When the absolute value of the frequency shift rate is higher than the third threshold, the first integration time is set. When the absolute value of the frequency shift rate is between the third threshold and the fourth threshold, the second integration time is set. When the absolute value of the frequency shift rate is lower than the fourth threshold, the third integration time is set. The third integration time is greater than the second integration time, and the second integration time is greater than the first integration time.

[0068] The set of capture parameters is obtained by setting the frequency of the numerically controlled oscillator, the frequency search window range, and the integration time.

[0069] Specifically, a predicted frequency value for a future moment is selected from the carrier frequency offset trajectory. This moment is typically chosen as a prediction point 50 milliseconds away from the current moment. This choice of time point comprehensively considers system processing latency and prediction accuracy. Processing latency includes the sum of signal acquisition time, feature extraction time, neural network inference time, and parameter loading time. Prediction accuracy decreases as prediction time increases. The 50-millisecond lead balances latency compensation and prediction reliability in engineering practice. Subtracting this predicted frequency value from the receiver's intermediate frequency (IF) yields the frequency setpoint for the numerically controlled oscillator (CNC). The IF is the carrier frequency after down-conversion of the signal. For StarLink signals, the IF is typically set to 70MHz, and for GPS signals, it's typically set to 4.092MHz. This subtraction operation performs frequency pre-compensation, allowing the local carrier to adjust to the signal's future frequency position in advance. Dividing the frequency setpoint by the system clock frequency yields the normalized frequency. The system clock frequency is the reference clock for the CNC oscillator, typically 100MHz. The normalized frequency represents the proportion of phase accumulation within each clock cycle. The phase increment is obtained by multiplying the normalized frequency by the maximum value corresponding to the phase accumulator bit width. The phase accumulator bit width determines the frequency resolution; the maximum value of a 48-bit phase accumulator is 2 to the power of 48, or 281474976710656. Using an integer phase increment facilitates hardware implementation. The phase increment value is written to the phase increment register of the numerically controlled oscillator. The phase accumulator reads this register value each clock cycle and adds it to the current accumulated phase. The high-order bits of the accumulated phase are used as an address index to look up the sine and cosine waveform table stored in ROM. The waveform table outputs the sine and cosine values ​​corresponding to the phase, forming the output signal of the numerically controlled oscillator.

[0070] When determining the reliability level of a prediction based on its uncertainty, a first threshold and a second threshold need to be set. The first threshold is set based on the prediction error statistics of the LSTM network on the validation set. A typical value is obtained by statistically analyzing the root mean square error (RMSE) of prediction points over 50 milliseconds. 0.5 times this typical value is used as the first threshold. When the prediction uncertainty is below this threshold, it indicates that the fluctuation range of the prediction result is less than half of the average error, and the prediction is highly reliable. The second threshold is set to 1.5 times the typical error. When the prediction uncertainty is above this threshold, it indicates that the prediction fluctuation range exceeds 1.5 times the average error, and the prediction reliability is low. For Starlink low-Earth orbit (LEO) satellite signals, the RMSE of the LSTM network on the validation set over 50 milliseconds is approximately 400 Hz, so the first threshold is set to 200 Hz and the second threshold is set to 600 Hz. For GPS medium-to-high Earth orbit (MEO) satellite signals, the prediction error on the validation set is approximately 50 Hz, so the first threshold is set to 25 Hz and the second threshold is set to 75 Hz. When the prediction uncertainty is below a first threshold, a first frequency search window range is set. This first window range takes into account prediction error and receiver local oscillator stability, and is calculated by multiplying the first threshold by two to obtain the width of a single-sided window. For StarLink signals, the first window range is ±400Hz, and for GPS signals, it is ±50Hz. When the prediction uncertainty is between the first and second thresholds, a second frequency search window range is set, which is 2.5 times the size of the first window range. For StarLink signals, the second window range is ±1000Hz, and for GPS signals, it is ±125Hz. When the prediction uncertainty is above the second threshold, a third frequency search window range is set, which is 5 times the size of the first window range. For StarLink signals, the third window range is ±2000Hz, and for GPS signals, it is ±250Hz. This progressively expanding window range ensures coverage of the signal's true frequency deviation at different reliability levels.

[0071] When determining the dynamic level of a signal based on its frequency shift rate, a third and fourth threshold need to be set. These thresholds are based on the Doppler frequency shift rate characteristics of satellites at different orbital altitudes. Low Earth orbit (LEO) satellites typically orbit at altitudes of 500 to 1200 kilometers, with a maximum radial velocity of approximately 7 kilometers per second relative to the ground receiver. This corresponds to a maximum Doppler frequency shift rate of approximately 150,000 Hz per second in the Ku-band. Medium Earth orbit (MEO) satellites orbit at an altitude of approximately 20,000 kilometers, with a maximum radial velocity of approximately 1 kilometer per second. This corresponds to a maximum Doppler frequency shift rate of approximately 5,000 Hz per second in the L-band. The third threshold is set to two-thirds of the typical LEO satellite frequency shift rate, i.e., 100,000 Hz per second. When the absolute value of the frequency shift rate exceeds this threshold, the signal is considered high dynamic. The fourth threshold is set to one-third of the typical LEO satellite frequency shift rate, i.e., 50,000 Hz per second. When the absolute value of the frequency shift rate is below this threshold, the signal is considered low dynamic. When the absolute value of the frequency shift rate is higher than the third threshold, the signal frequency changes rapidly. A first integration time of 5 milliseconds is set. This short integration time avoids frequency shifts during integration that could lead to correlation peak broadening and energy dispersion. Within 5 milliseconds, the signal frequency change at a frequency shift rate of 100,000 Hz is 500 Hz, which has a limited impact relative to the signal bandwidth. When the absolute value of the frequency shift rate is between the third and fourth thresholds, the signal dynamics are moderate. A second integration time of 10 milliseconds is set. This duration strikes a balance between energy accumulation and frequency shifts. Within 10 milliseconds, the signal frequency change at a frequency shift rate of 75,000 Hz is 750 Hz. When the absolute value of the frequency shift rate is lower than the fourth threshold, the signal dynamics are slow. A third integration time of 20 milliseconds is set. This long integration time fully accumulates signal energy and improves processing gain. Within 20 milliseconds, the signal frequency change at a frequency shift rate of 25,000 Hz is 500 Hz, with minimal impact from correlation peak broadening. The difference in processing gain among the three integration times is reflected in the logarithm of the integration time ratio. The processing gain of the third integration time relative to the first integration time is increased by 10 times logarithmically, or 6 dB. This gain improvement significantly enhances acquisition performance under low signal-to-noise ratio conditions. The adjusted numerically controlled oscillator frequency setpoint, the center frequency and upper and lower limit frequencies of the frequency search window, the frequency step value, and the integration time parameters are organized into a set of acquisition parameters according to a fixed format. This set of parameters is transmitted to the acquisition execution module through a register interface or a bus interface. The acquisition module configures the correlator, accumulator, and decision unit based on these parameters to complete signal acquisition.

[0072] Figure 2 This is a schematic diagram showing the frequency search window range under different reliability levels in the embodiments of this application. Figure 2The figure illustrates a strategy for dynamically adjusting the frequency search window range based on prediction uncertainty. The horizontal axis represents the prediction reliability levels, categorized into high reliability (below the first threshold), medium reliability (between the first and second thresholds), and low reliability (above the second threshold). The vertical axis represents the search window range in ±Hz. As can be seen from the figure, for Starlink low-Earth orbit satellites, due to their large Doppler shift rate, the search window range gradually expands from ±400Hz for high reliability to ±1000Hz for medium reliability and ±2000Hz for low reliability. Conversely, for GPS medium- and high-Earth orbit satellites, due to their smaller shift rate, the search window range is correspondingly narrower, expanding from ±50Hz for high reliability to ±125Hz for medium reliability and ±250Hz for low reliability. This hierarchical setting strategy balances acquisition success rate with computational efficiency.

[0073] In one specific embodiment, step S5 includes:

[0074] The classification result of the interference type determines whether it belongs to rain attenuation interference, adjacent channel interference, or malicious interference;

[0075] When the judgment result is rain attenuation interference, calculate the signal-to-noise ratio decrease rate. When the signal-to-noise ratio decrease rate exceeds the rain attenuation judgment threshold, extend the integration time, query satellite ephemeris data to calculate the elevation angle of the currently tracked satellite, and when the elevation angle is lower than the elevation angle threshold, search for high elevation angle candidate satellites in the visible satellite list and send a switching command.

[0076] When the judgment result is adjacent channel interference, peak detection is performed on the power spectral density curve of the time-frequency two-dimensional matrix, the frequency points that exceed the interference detection threshold are recorded, the interference frequency point set is obtained, a notch filter is designed for the interference frequency point set and the filter transfer function is generated, and the coefficients of the filter transfer function are loaded into the configurable filter module.

[0077] When the judgment result is malicious interference, the instantaneous drop in the current signal-to-noise ratio is calculated. When the instantaneous drop exceeds the malicious interference judgment threshold, the frequency band configuration table of different constellations is queried to select the alternative constellation, and frequency switching command and signal system switching command are sent to the radio frequency front end.

[0078] Specifically, the system determines whether the interference type classification result belongs to rain attenuation interference, adjacent channel interference, or malicious interference. When the category label is rain attenuation interference, the signal-to-noise ratio (SNR) values ​​at the current time and multiple times over a past period are read. The SNR is the ratio of signal power to noise power. Signal power is estimated using relevant peak values, and noise power is estimated using the variance of noise samples around the relevant peak values. The SNR values ​​at each time point are arranged chronologically, and a linear fit is performed on the time and SNR data. The slope of the fitted line is the SNR decrease rate; a negative slope indicates a decrease in SNR, and a large absolute value of the slope indicates a rapid decrease. When the SNR decrease rate exceeds the rain attenuation threshold, rain attenuation interference is confirmed. The integration time is extended by a fixed percentage from the current set value. Extending the integration time increases signal energy accumulation to compensate for rain attenuation loss. Satellite ephemeris data is queried to obtain the orbital parameters of the tracking satellite. Based on the latitude, longitude, and altitude of the receiver position, the geocentric coordinates of the satellite position are calculated. The angle between the vector pointing from the receiver to the satellite and the ground plane is the satellite elevation angle. When the elevation angle is lower than the set elevation angle threshold, the system searches the currently visible satellite list for satellites with elevation angles higher than the threshold. The visible satellite list is calculated in real time based on ephemeris forecasts and receiver position. The satellite with the highest elevation angle is selected as the candidate satellite, and a switching command is sent to the radio frequency front end. When the category label is adjacent channel interference, the power spectral density curve is extracted along the frequency axis of the time-frequency two-dimensional matrix. The power spectral density curve describes the distribution of signal power on the frequency axis. The average power of all frequency points on the curve is calculated. The interference detection threshold is set as the average value plus a fixed margin. The power spectral density curve is traversed to find all frequency points where the power value exceeds the threshold. These frequency points constitute the interference frequency point set. For each frequency in the interference frequency set, a notch filter is designed. The normalized angular frequency is calculated by dividing the interference frequency by the sampling frequency and then multiplying by 2 times pi. The notch depth control parameter is selected to be close to 1. Based on the normalized angular frequency and control parameters, the numerator and denominator coefficients of the filter transfer function are calculated. The numerator coefficient includes a cosine term related to the normalized angular frequency, and the denominator coefficient includes the product of the control parameter and the cosine term. The coefficients are loaded into the register of the filter module, and the filter filters the baseband signal in real time. When the category label is malicious interference, the difference between the current signal-to-noise ratio and the previous signal-to-noise ratio is calculated to obtain the instantaneous drop amplitude. When the instantaneous drop amplitude exceeds the malicious interference judgment threshold, malicious interference is confirmed. The constellation frequency band configuration table is queried to obtain the operating frequency band of each constellation. The difference between the current frequency band and the frequency bands of other constellations is calculated, and the constellation with the largest difference is selected as the candidate constellation. A frequency switching command is sent to the RF front end, which includes the new local oscillator frequency. At the same time, a signal mode switching command is sent, which includes the new modulation mode and frame structure parameters.

[0079] In one specific embodiment, step S6 includes:

[0080] The ratio of the signal correlation peak power to the noise floor power is periodically calculated, the carrier-to-noise ratio is calculated based on the ratio, the number of errors in the pilot symbol hard decision is counted, the bit error rate is calculated based on the number of errors, and the time interval from successful acquisition to loss of lock is recorded to obtain the lock holding time.

[0081] Compare the currently measured carrier-to-noise ratio, bit error rate, and lock-and-hold time with the baseline values, calculate the ratio of the change in each indicator to the baseline value, and obtain the performance improvement rate.

[0082] Determine if the performance improvement rate is lower than the set threshold. If the determination result is yes and the duration exceeds the determination time, trigger the policy upgrade and select the corresponding upgrade policy according to the interference type classification result.

[0083] Sample data whose accuracy exceeds the threshold within multiple consecutive evaluation periods are stored in a local sample buffer. When the local sample buffer reaches its capacity limit, the buffer samples are mixed with the original training set, and the weights of the task branch layers of the multi-task neural network are fine-tuned to obtain the updated network parameters.

[0084] Specifically, signal quality monitoring is performed periodically, and three indicators—carrier-to-noise ratio (CNR), bit error rate (BER), and lock-on time—are calculated within each monitoring period. When calculating the CNR, correlation operations are performed on the received signal. These operations multiply the received signal point-by-point with a locally generated pseudocode sequence and accumulate the results to find the correlation peak. The square of the amplitude of the correlation peak yields the signal correlation power. Several non-peak points are selected around the correlation peak, and the power variance of these points is calculated to estimate the noise power. The signal correlation power is subtracted from the noise power, and the result is divided by the noise power to obtain the signal-to-noise ratio (SNR). The logarithm of the SNR is multiplied by 10 to obtain the decibel value, which is then added to the logarithm of the equivalent noise bandwidth and multiplied by 10 to obtain the CNR. The BER calculation utilizes pilot symbols in the signal. The content of the pilot symbols is a pre-agreed known sequence between the transmitter and receiver. The receiver demodulates the pilot symbols, mapping each symbol point to the nearest standard symbol based on its distance from the decision threshold. The demodulated symbol sequence is then compared symbol-by-symbol with the known pilot symbol sequence, and the number of inconsistent symbols is counted. The number of inconsistent symbols is divided by the total number of pilot symbols to obtain the BER. The lock-and-hold time begins counting from the moment the signal is successfully acquired and enters the tracking state. The tracking loop continuously monitors relevant peak values ​​and carrier phase. When the relevant peak value falls below the lock-in threshold or the carrier phase error exceeds the threshold, a lock-out is determined. The lock-and-hold time is obtained by subtracting the acquisition success time from the lock-out time. The carrier-to-noise ratio (CNR), bit error rate (BER), and lock-and-hold time measured in the current monitoring cycle are compared with the baseline values ​​recorded before the anti-interference strategy was implemented. The baseline value is the average of several cycles before the anti-interference strategy was implemented. Subtracting the baseline value from the current value yields the change. Dividing the change by the baseline value yields the improvement rate of the indicator. A positive CNR improvement rate indicates an increase in CNR, a negative BER improvement rate indicates a decrease in BER (i.e., performance improvement), and a positive lock-and-hold time improvement rate indicates an extension of the lock-in time. The overall performance improvement rate is calculated by combining the improvement rates of the three indicators. It is then determined whether the performance improvement rate is lower than a set threshold and whether the duration of this state exceeds the determination time. A duration counter records the number of cycles in which the improvement rate is continuously lower than the threshold. When the counter exceeds a set value, a strategy upgrade is triggered. Based on the interference type classification results, an upgrade strategy is selected. For rain-attenuation interference, the upgrade strategy extends the integration time beyond the current value and sends a code rate reduction command to the encoding module. The encoding module then switches to a lower code rate encoding scheme. For adjacent-channel interference, the upgrade strategy increases the number of notch filter cascades by connecting new notch units in series after the existing notch filters. For malicious interference, the upgrade strategy switches from the current constellation to a secondary alternative constellation and queries the constellation priority list to select the next priority constellation. Sample data from multiple consecutive evaluation periods are filtered, and the recognition accuracy is calculated as the number of correctly identified samples divided by the total number of samples. When the accuracy exceeds a set threshold, the sample data for that period is stored in a local buffer. The sample data includes feature vectors, true class labels, and network predicted labels.The local sample buffer has a capacity limit. When the number of samples stored in the buffer reaches the limit, model fine-tuning is triggered. Samples are proportionally drawn from the original training set, and these samples are mixed with the new samples in the buffer to form a fine-tuned training set. When fine-tuning a multi-task neural network, the weight parameters of the shared feature extraction layer are frozen, and only the weights of the task branch layers are allowed to be updated. The fine-tuned training set is input into the network, and the loss function between the network output and the true label is calculated. The gradient of the loss function with respect to the weights of the task branch layers is calculated using the backpropagation algorithm. The weight parameters are updated based on the gradient and the learning rate. After multiple iterations, the updated network parameters are obtained.

[0085] The above describes the intelligent identification and adaptive acquisition method for multi-constellation satellite signals in the embodiments of this application. The following describes the intelligent identification and adaptive acquisition system for multi-constellation satellite signals in the embodiments of this application. Please refer to [link to relevant documentation]. Figure 3 One embodiment of the multi-constellation satellite signal intelligent identification and adaptive acquisition system in this application includes:

[0086] The transformation module is used to perform short-time Fourier transform on the received signal, extract the frequency-time trajectory matrix, calculate the Doppler frequency shift rate and frequency shift acceleration, and combine the modulation domain features to form a feature vector;

[0087] The input module is used to input the feature vector into the multi-task neural network, and output constellation category recognition results and interference type classification results through a shared feature extraction layer and a dual-branch task head;

[0088] The prediction module is used to predict the carrier frequency offset trajectory within a future time window based on historical frequency sequences and a long short-term memory network when low-Earth orbit satellite signals are identified, while outputting the prediction uncertainty.

[0089] The adjustment module is used to adjust the frequency of the numerically controlled oscillator in advance according to the carrier frequency offset trajectory, dynamically set the frequency search window range according to the prediction uncertainty, and adaptively adjust the integration time according to the frequency shift rate.

[0090] The switching module is used to execute corresponding anti-interference strategies based on the interference type classification results, extend the integration time and switch to high elevation angle satellites for rain attenuation interference, configure notch filters for adjacent channel interference, and perform constellation frequency band switching for malicious interference.

[0091] The upgrade module is used to periodically monitor the carrier-to-noise ratio, bit error rate, and lockout duration. When the performance improvement rate is lower than a set threshold, a policy upgrade is triggered.

[0092] above Figure 3The intelligent identification and adaptive acquisition system for multi-constellation satellite signals in this embodiment of the invention is described in detail from the perspective of modular functional entities. The intelligent identification and adaptive acquisition device for multi-constellation satellite signals in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0093] Reference Figure 4 This invention also provides a multi-constellation satellite signal intelligent identification and adaptive acquisition device, which can be a server, and its internal structure can be as follows: Figure 4 As shown, the multi-constellation satellite signal intelligent identification and adaptive acquisition device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computing and control capabilities. The memory of the multi-constellation satellite signal intelligent identification and adaptive acquisition device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the multi-constellation satellite signal intelligent identification and adaptive acquisition device stores the data corresponding to this embodiment. The network interface of the multi-constellation satellite signal intelligent identification and adaptive acquisition device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0094] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the multi-constellation satellite signal intelligent identification and adaptive acquisition device to which the present invention is applied.

[0095] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the multi-constellation satellite signal intelligent identification and adaptive acquisition method.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0097] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a multi-constellation satellite signal intelligent identification and adaptive acquisition device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0098] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-constellation satellite signal intelligent recognition and adaptive acquisition method, characterized in that, The method comprises: S1 step: short-time Fourier transform is performed on the received signal, a frequency-time trajectory matrix is extracted, Doppler shift rate and frequency shift acceleration are calculated, and modulation domain features are combined to form a feature vector; S2 step: inputting the feature vector into a multi-task neural network, outputting constellation category recognition results and interference type classification results through a shared feature extraction layer and a double-branch task head; S3 step: when a low-orbit satellite signal is recognized, based on a historical frequency sequence, a long short-term memory network is used to predict a carrier frequency offset trajectory in a future time window, and a prediction uncertainty is outputted simultaneously; S4 step: adjusting a numerically controlled oscillator frequency in advance according to the carrier frequency offset trajectory, dynamically setting a frequency search window range according to the prediction uncertainty, and adaptively adjusting an integration time according to the frequency shift rate; S5 step: executing a corresponding anti-interference strategy according to the interference type classification result, prolonging the integration time and switching to a high-elevation satellite for rain attenuation interference, configuring a notch filter for adjacent frequency interference, and performing constellation band switching for malicious interference; S6 step: periodically monitoring a carrier-to-noise ratio, a bit error rate and a lock holding time, and triggering a strategy upgrade when a performance improvement rate is lower than a set threshold. 2.The method of claim 1, wherein, The S1 step comprises: performing digital down-conversion processing on the received signal to obtain baseband complex signals of in-phase components and quadrature components; performing short-time Fourier transform on the baseband complex signals to obtain a time-frequency two-dimensional matrix; searching for a spectral energy peak value point in each time slice of the time-frequency two-dimensional matrix, recording a peak frequency, and obtaining a frequency sequence; performing first-order difference operation on the frequency sequence and dividing by a time interval to obtain a Doppler shift rate, and performing second-order difference operation on the Doppler shift rate and dividing by a time interval to obtain a frequency shift acceleration; performing fourth power operation on the baseband complex signals and then performing fast Fourier transform, detecting a peak strength at four times the carrier frequency, and obtaining modulation domain features; obtaining a feature vector from the Doppler shift rate, the frequency shift acceleration and the modulation domain features. 3.The method of claim 2, wherein, The S2 step comprises: inputting the feature vector into a one-dimensional convolutional neural network layer of a multi-task neural network for feature extraction, obtaining a shared feature vector through multi-layer convolution operation and activation function processing; inputting the shared feature vector into a global average pooling layer for dimension reduction processing to obtain a fixed-dimension feature representation; simultaneously inputting the fixed-dimension feature representation into a constellation recognition task branch and an interference identification task branch, the constellation recognition task branch comprising a constellation fully connected layer and a constellation Softmax layer, outputting a constellation category probability vector, and the interference identification task branch comprising an interference fully connected layer and an interference Softmax layer, outputting an interference type probability vector; comparing probability values of components in the constellation category probability vector, selecting a category label corresponding to a maximum probability value to obtain a constellation category recognition result, and comparing probability values of components in the interference type probability vector, selecting a category label corresponding to a maximum probability value to obtain an interference type classification result.

4. The method of claim 3, wherein the plurality of constellations are determined based on a plurality of satellite systems. The S3 step comprises: judging whether the constellation category recognition result is a low earth orbit satellite signal, and when the result is yes, extracting a historical frequency sequence in the frequency sequence; performing normalization processing on the historical frequency sequence, calculating a mean value and a standard deviation, subtracting each frequency value in the historical frequency sequence from the mean value and dividing by the standard deviation to obtain a normalized frequency sequence; inputting the normalized frequency sequence into a long short-term memory network, updating cell state and hidden state through gate operation of a forget gate, an input gate and an output gate, inputting the normalized frequency sequence into a fully connected layer after weighting by an attention mechanism layer to obtain a normalized prediction sequence; performing inverse normalization processing on the normalized prediction sequence, multiplying each prediction value in the normalized prediction sequence by the standard deviation and adding the mean value to obtain a carrier frequency offset trajectory, calculating a standard deviation of the carrier frequency offset trajectory through multiple forward propagations to obtain a prediction uncertainty.

5. The method of claim 4, wherein, The S4 step comprises: extracting a predicted frequency value at a future time according to the carrier frequency offset trajectory, subtracting the predicted frequency value from an intermediate frequency frequency to obtain a numerically controlled oscillator frequency setting value, multiplying the numerically controlled oscillator frequency setting value by a value corresponding to a phase accumulator bit width after dividing by a system clock frequency to obtain a phase increment, and adjusting a phase accumulator of the numerically controlled oscillator according to the phase increment; judging a prediction reliability level according to the prediction uncertainty, setting a first frequency search window range when the prediction uncertainty is lower than a first threshold value, setting a second frequency search window range when the prediction uncertainty is between the first threshold value and a second threshold value, and setting a third frequency search window range when the prediction uncertainty is higher than the second threshold value, the third frequency search window range being greater than the second frequency search window range, and the second frequency search window range being greater than the first frequency search window range; judging a signal dynamic level according to the frequency shift rate, setting a first integration time when an absolute value of the frequency shift rate is higher than a third threshold value, setting a second integration time when the absolute value of the frequency shift rate is between the third threshold value and a fourth threshold value, and setting a third integration time when the absolute value of the frequency shift rate is lower than the fourth threshold value, the third integration time being greater than the second integration time, and the second integration time being greater than the first integration time; obtaining a capture parameter set by using the numerically controlled oscillator frequency setting value, the frequency search window range and the integration time.

6. The method of claim 5, wherein the step of identifying the satellite signal comprises: The S5 step comprises: judging whether the interference type classification result belongs to rain fade interference, adjacent frequency interference or malicious interference; when the result is rain fade interference, calculating a signal-to-noise ratio drop rate, lengthening the integration time when the signal-to-noise ratio drop rate exceeds a rain fade determination threshold value, querying satellite ephemeris data to calculate an elevation angle of a currently tracked satellite, searching for a high-elevation angle candidate satellite in a visible satellite list and sending a switching instruction when the elevation angle is lower than an elevation angle threshold value. When the judgment result is adjacent frequency interference, peak value detection is performed on the power spectrum density curve of the time-frequency two-dimensional matrix, the frequency point positions exceeding the interference detection threshold are recorded to obtain an interference frequency point set, a notch filter is designed for the interference frequency point set and a filter transfer function is generated, and coefficients of the filter transfer function are loaded into a configurable filter module; When the judgment result is malicious interference, an instantaneous drop amplitude of a current signal-to-noise ratio is calculated, when the instantaneous drop amplitude exceeds a malicious interference judgment threshold, a candidate constellation is selected by querying a frequency band configuration table of different constellations, and a frequency switching instruction and a signal system switching instruction are sent to a radio frequency front end.

7. The method of claim 1, wherein, The S6 step comprises: Periodically calculating a ratio of a signal correlation peak power to a noise floor power, calculating a carrier-to-noise ratio according to the ratio, calculating an error code rate according to a number of errors of a pilot symbol hard decision, and recording a time interval from successful acquisition to lock loss to obtain a lock holding time; Comparing the currently measured carrier-to-noise ratio, error code rate and lock holding time with reference values, calculating a ratio of a change amount of each index to the reference value to obtain a performance improvement rate; Judging whether the performance improvement rate is lower than a set threshold, triggering a strategy upgrade when the judgment result is yes and a duration exceeds a judgment length, and selecting a corresponding upgrade strategy according to the interference type classification result; Storing sample data with an accuracy rate exceeding a threshold in a local sample buffer for a plurality of evaluation periods, mixing buffer samples with an original training set when the local sample buffer reaches an upper limit of a capacity, fine-tuning task branch layer weights of the multi-task neural network to obtain updated network parameters.

8. A multi-constellation satellite signal intelligent identification and adaptive acquisition system, characterized in that, The multi-constellation satellite signal intelligent recognition and adaptive acquisition system comprises: A transformation module configured to perform short-time Fourier transform on a received signal, extract a frequency-time trajectory matrix, calculate a Doppler frequency shift rate and a frequency shift acceleration, and combine modulation domain features to form a feature vector; An input module configured to input the feature vector into a multi-task neural network, output constellation category recognition results and interference type classification results through a shared feature extraction layer and a double-branch task head; A prediction module configured to predict a carrier frequency offset trajectory in a future time window through a long short-term memory network based on a historical frequency sequence when a low-orbit satellite signal is recognized, and simultaneously output a prediction uncertainty; An adjustment module configured to adjust a number-controlled oscillator frequency in advance according to the carrier frequency offset trajectory, dynamically set a frequency search window range according to the prediction uncertainty, and adaptively adjust an integration time according to the frequency shift rate; A switching module configured to execute corresponding anti-interference strategies according to the interference type classification results, extend the integration time and switch to a high-elevation-angle satellite for rain attenuation interference, configure a notch filter for adjacent frequency interference, and execute constellation frequency band switching for malicious interference; An upgrade module configured to periodically monitor a carrier-to-noise ratio, an error code rate and a lock holding time, and trigger a strategy upgrade when a performance improvement rate is lower than a set threshold.

9. A multi-constellation satellite signal intelligent identification and adaptive acquisition device, characterized in that, The computer program is stored in the memory and comprises computer program elements for performing the steps of the method when the computer program is run on the processor.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when run on the processor, causes the processor to perform the steps of the method of any one of claims 1 to 7.

Images