Bit timing estimation method and device suitable for satellite communication

Abstract

The present invention discloses a method and device for estimating bit timing suitable for satellite communication. The method includes calculating the time correlation parameter and signal-to-noise ratio parameter of the satellite signal, determining the time window size; constructing a window feature sequence based on a bidirectional recursive processing unit, and performing window segment sampling on the received signal; extracting the autocorrelation function of each segment of the signal to obtain a time domain feature vector, a frequency domain feature vector and a phase feature vector; performing three-level nonlinear mapping on the three types of feature vectors to obtain an adaptive weight vector; using the weight vector to perform weighted fusion on the feature and perform cubic spline interpolation, calculating the error compensation value in combination with the channel state parameter, and obtaining a fine timing estimation value; extracting synchronization features according to the timing estimation value and solving the optimization problem in segments to obtain time synchronization parameters; calculating the fused feature vector, obtaining a decision result through pattern recognition and parameter optimization, and completing the bit timing estimation. The present invention improves the timing accuracy and reliability of the satellite communication system.

Description

Technical Field

[0001] The present invention belongs to the field of satellite communication, and in particular to a method and a device for estimating bit timing suitable for satellite communication. Background Art

[0002] In satellite communication systems, accurate bit timing estimation is crucial to achieve reliable data transmission and system synchronization. Bit timing estimation faces severe technical challenges due to the high altitude of satellite orbits far from the ground, long signal propagation paths, and the influence of multiple factors such as ionospheric scintillation, Doppler shift, and atmospheric refraction. Especially in low-orbit satellite communication systems, due to the high-speed movement of satellites relative to ground stations, channel characteristics change rapidly, which makes accurate bit timing estimation a key technical link to ensure system performance.

[0003] At present, the commonly used bit timing estimation methods mainly include the correlation peak detection method based on the maximum likelihood criterion, the interpolation algorithm based on oversampling, and the timing estimation method based on cyclic prefix. The correlation peak detection method extracts timing information by designing a matched filter, but it is easy to generate false peaks under low signal-to-noise ratio conditions; although the interpolation algorithm can improve the sampling accuracy, the computational complexity is high and it is difficult to meet the real-time processing requirements; the method based on cyclic prefix uses the periodic characteristics of the signal for timing estimation, but it is more sensitive to fast channel fading and its performance is easily degraded. These traditional methods have limitations to varying degrees when dealing with timing estimation problems in complex spatial channel environments.

[0004] In-depth analysis found that the current timing estimation technology still has the following specific problems: First, the existing feature extraction methods often use a fixed window size for processing, which is difficult to adapt to the dynamic changes of channel conditions, resulting in insufficient feature expression capabilities; second, in the process of multi-feature fusion, the determination of feature weights lacks an adaptive mechanism, and the contribution of each feature cannot be adjusted in real time according to the channel state; third, when processing non-uniformly sampled data, traditional interpolation algorithms are prone to overshoot and oscillation at the interpolation point, affecting the timing accuracy; fourth, the error compensation model fails to fully consider the nonlinear relationship between channel state parameters and timing errors, and the compensation effect is not ideal; finally, the existing decision-making mechanism lacks a reliability evaluation system, making it difficult to accurately quantify the confidence level of the timing results, reducing the overall reliability of the system. These technical problems seriously restrict the synchronization performance and communication quality of satellite communication systems in complex space environments. Summary of the invention

[0005] The purpose of the invention is to provide a method and device for bit timing estimation suitable for satellite communications to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution is applicable to a bit timing estimation method for satellite communications, comprising the following steps:

[0007] S1, obtaining an original complex baseband signal, mapping the original complex baseband signal to a four-dimensional feature space, and obtaining a feature matrix; performing piecewise nonlinear transformation and dimensionality reduction processing on the feature matrix to obtain a transformed feature matrix; using a recursive sliding window to process the transformed feature matrix, and extracting a time synchronization feature vector; based on the time synchronization feature vector, performing a multi-layer adaptive fusion operation to obtain an initial timing point estimate;

[0008] S2. Based on the initial timing point estimate, extract the received signal of the corresponding time window, and input it into the adaptive filter group for processing to obtain the time domain feature vector, frequency domain feature vector and phase feature vector; process the time domain feature vector, frequency domain feature vector and phase feature vector through a nonlinear mapping network to obtain an adaptive weight vector; based on the adaptive weight vector, perform fusion calculation and interpolation optimization on the time domain feature vector, frequency domain feature vector and phase feature vector to obtain a fine timing estimate; based on the fine timing estimate, construct a state space model, perform dynamic tracking, and obtain the final timing correction value;

[0009] S3, performing multi-dimensional performance evaluation on the final timing correction value and the pre-stored historical timing sequence to obtain a performance index vector; based on the performance index vector, performing anti-interference processing on the received signal to obtain an anti-interference processed signal; obtaining current system parameters, performing optimization calculation based on the current system parameters and the performance index vector to obtain an optimized parameter set;

[0010] S4. Based on the anti-interference processing signal and the optimization parameter set, perform time-frequency joint analysis and phase space reconstruction to obtain a fused feature vector; based on the fused feature vector, perform multi-level decision operations to obtain a decision result; based on the decision result and pre-stored historical performance data, perform system self-calibration to obtain a calibration parameter set; input the calibration parameter set and the pre-stored historical data into the prediction optimization module to obtain the final optimization parameters and performance prediction report.

[0011] A bit timing estimation device suitable for satellite communications, comprising:

[0012] at least one processor; and,

[0013] a memory communicatively connected to at least one of the processors; wherein,

[0014] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the bit timing estimation method applicable to satellite communication.

[0015] Beneficial effects: The present invention provides a comprehensive characterization of signal characteristics, ensuring the accuracy of the estimation results; through multi-dimensional performance evaluation and real-time monitoring, it ensures the stable operation of the system, improves the synchronization performance of the satellite communication system, and achieves high accuracy and high reliability of satellite communication median timing estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Flow chart of the method of the present invention.

[0017] Figure 2 This is a flow chart of step S1 of the present invention.

[0018] Figure 3 This is a flow chart of step S2 of the present invention.

[0019] Figure 4 This is a flow chart of step S3 of the present invention.

[0020] Figure 5 This is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0021] like Figure 1 As shown, the present application proposes a bit timing estimation method suitable for satellite communication, comprising the following steps:

[0022] S1, obtaining an original complex baseband signal, mapping the original complex baseband signal to a four-dimensional feature space, and obtaining a feature matrix; performing piecewise nonlinear transformation and dimensionality reduction processing on the feature matrix to obtain a transformed feature matrix; using a recursive sliding window to process the transformed feature matrix, and extracting a time synchronization feature vector; based on the time synchronization feature vector, performing a multi-layer adaptive fusion operation to obtain an initial timing point estimate;

[0023] S2. Based on the initial timing point estimate, extract the received signal of the corresponding time window, and input it into the adaptive filter group for processing to obtain the time domain feature vector, frequency domain feature vector and phase feature vector; process the time domain feature vector, frequency domain feature vector and phase feature vector through a nonlinear mapping network to obtain an adaptive weight vector; based on the adaptive weight vector, perform fusion calculation and interpolation optimization on the time domain feature vector, frequency domain feature vector and phase feature vector to obtain a fine timing estimate; based on the fine timing estimate, construct a state space model, perform dynamic tracking, and obtain the final timing correction value;

[0024] S3, performing multi-dimensional performance evaluation on the final timing correction value and the pre-stored historical timing sequence to obtain a performance index vector; based on the performance index vector, performing anti-interference processing on the received signal to obtain an anti-interference processed signal; obtaining current system parameters, performing optimization calculation based on the current system parameters and the performance index vector to obtain an optimized parameter set;

[0025] S4. Based on the anti-interference processing signal and the optimization parameter set, perform time-frequency joint analysis and phase space reconstruction to obtain a fused feature vector; based on the fused feature vector, perform multi-level decision operations to obtain a decision result; based on the decision result and pre-stored historical performance data, perform system self-calibration to obtain a calibration parameter set; input the calibration parameter set and the pre-stored historical data into the prediction optimization module to obtain the final optimization parameters and performance prediction report.

[0026] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0027] S11, obtaining an original complex baseband signal, processing the original complex baseband signal through an amplitude feature extraction function to obtain a signal amplitude feature vector; processing the original complex baseband signal through a phase feature extraction function to obtain a phase jump feature vector; processing the original complex baseband signal through a frequency feature extraction function to obtain a frequency offset feature vector; processing the original complex baseband signal through a time feature extraction function to obtain a time-related feature vector; combining the signal amplitude feature vector, the phase jump feature vector, the frequency offset feature vector and the time-related feature vector to obtain a feature matrix;

[0028] S12, calculating the norm value of the characteristic matrix to obtain a norm determination result; based on the norm determination result, selecting a corresponding nonlinear transformation function; based on the nonlinear transformation function, processing the characteristic matrix to obtain a transformed characteristic matrix;

[0029] S13, based on the pre-stored signal-to-noise ratio parameter and time correlation parameter, calculate the window size value; based on the window size value, recursively slide the window processing on the transformed feature matrix to obtain a window feature sequence; calculate the weighted feature combination value, the autocorrelation function envelope peak position and the feature difference of each window in the window feature sequence; combine the weighted feature combination value, the autocorrelation function envelope peak position and the feature difference to obtain a time synchronization feature vector;

[0030] S14. Perform an exponentially weighted moving average filter on the time synchronization feature vector to obtain a first-level filtering result; multiply the first-level filtering result by a preset adaptive weight matrix to obtain a feature weight; perform a weighted operation on the feature weight and the first-level filtering result, take the time point corresponding to the maximum value, and obtain an initial timing point estimate.

[0031] In one embodiment of the present application, the received original complex baseband signal sequence r(t) is mapped to a four-dimensional feature space H, where each dimension represents: h1: signal amplitude feature vector; h2: phase jump feature vector; h3: frequency offset feature vector; h4: time correlation feature vector; and the output feature matrix H = [h1, h2, h3, h4]. Based on the feature matrix H, a new piecewise nonlinear transformation function Φ(H) is applied: Φ(H) = α·tanh(β·H) + γ·sigmoid(Δ·H), when ||H||≤T1; Φ(H) = λ·exp(-μ·H) + η·sin(θ·H), when T1<||H||≤T2; Φ(H) = ρ·log(1 + σ·H), when ||H||>T2; where α, β, γ, Δ, λ, μ, η, θ, ρ and σ are adaptive parameters; T1 and T2 are dynamic thresholds; ||H|| represents the norm of the matrix H; and the transformed feature matrix H' is output.

[0032] Based on the transformed feature matrix H', a new recursive sliding window operator Ω is constructed: the window size W is dynamically adjusted: W = f(SNR, time correlation), where f( ) is a function and SNR is the signal-to-noise ratio; for each window, P1 = Ω(H'[t:t+W]) =∑(weight vector × nonlinear feature combination); P2 = envelope peak position of the autocorrelation function; P3 = feature difference between adjacent windows; output time synchronization feature vector TS = [P1, P2, P3]. Based on the time synchronization feature vector TS, an adaptive fusion estimator is applied: the first layer uses an improved EWMA (exponential weighted moving average) filter: Y1=λ1·TS+(1-λ1)·Y1_prev, where Y1 is the filtering result at the current moment, λ1 is the smoothing factor, and Y1_prev is the filtering result at the previous moment; the second layer performs nonlinear feature weight calculation: W = softmax(Y1 * A), where A is the adaptive weight matrix; the third layer performs the final estimation value calculation: τ_init = argmax(W·Y1); and outputs the initial timing point estimation value τ_init.

[0033] This embodiment realizes multi-dimensional extraction and optimization of signal features by mapping the complex baseband signal to a four-dimensional feature space and performing nonlinear transformation and dimensionality reduction processing. Specifically, by extracting the features of the four dimensions of amplitude, phase, frequency and time respectively, a complete feature representation space is constructed, so that the key features of the signal are fully captured. By adopting the combination of norm judgment and adaptive nonlinear transformation, the optimal transformation function can be selected according to the different signal features, thereby improving the adaptability and robustness of feature extraction. The introduction of the recursive sliding window realizes the dynamic capture of time-varying features. By calculating the weighted combination value of the window features, the peak value of the autocorrelation function envelope and the feature difference, the periodicity and mutation characteristics of the signal can be effectively identified. Finally, the exponentially weighted moving average filtering and adaptive weight fusion are adopted to not only suppress noise interference, but also realize the optimal combination of features, thereby obtaining a more accurate initial timing point estimation. This embodiment constructs an adaptive multi-dimensional feature extraction framework, and improves the accuracy and reliability of timing estimation by means of feature space mapping, nonlinear transformation and multi-layer fusion.

[0034] According to one aspect of the present application, step S11 is further:

[0035] S111, obtaining an original complex baseband signal, performing adaptive piecewise linear quantization processing, and obtaining a normalized signal; performing a multi-scale wavelet transform on the normalized signal to obtain a multi-scale coefficient matrix; based on the multi-scale coefficient matrix, calculating the energy distribution of each scale to obtain an energy distribution vector; inputting the energy distribution vector into a chaotic mapping unit for feature enhancement to obtain a signal amplitude feature vector;

[0036] S112, the original complex baseband signal is subjected to Hilbert transformation to obtain an analytical signal; the instantaneous phase of the analytical signal is calculated to obtain a phase sequence; the phase sequence is subjected to adaptive threshold detection to obtain a phase jump point; based on the phase jump point, a local phase change feature is extracted to obtain a local phase feature; the local phase feature is compressed through a neural network encoder to obtain a phase jump feature vector;

[0037] S113, performing segmented windowing processing on the original complex baseband signal to obtain a first segmented signal sequence; performing adaptive fractional Fourier transform on the first segmented signal sequence to obtain a spectrum sequence; calculating the center of gravity drift of the spectrum sequence to obtain a frequency drift sequence; performing fuzzy clustering processing on the frequency drift sequence to obtain a frequency offset feature vector;

[0038] S114, based on the original complex baseband signal, perform autocorrelation calculation to obtain an autocorrelation sequence; perform nonlinear feature extraction on the autocorrelation sequence to obtain a nonlinear feature sequence; calculate the recursive quantization entropy of the nonlinear feature sequence to obtain an entropy value sequence; process the entropy value sequence through a deep residual network to obtain a time-related feature vector;

[0039] S115. Based on the signal amplitude feature vector, the phase jump feature vector, the frequency offset feature vector and the time correlation feature vector, feature fusion is performed to obtain an initial feature combination; adaptive weight calculation is performed on the initial feature combination to obtain a feature weight vector; based on the feature weight vector, weighted combination is performed on the initial feature combination to obtain a feature matrix.

[0040] This embodiment achieves comprehensive capture and optimized characterization of signal features by constructing a multi-dimensional feature extraction system. The combination of adaptive piecewise linear quantization and multi-scale wavelet transform provides multi-resolution analysis capabilities for signal amplitude features, and the introduction of chaotic mapping enhances the distinguishability of features. Hilbert transform and adaptive threshold detection achieve precise positioning of phase jumps, and the application of neural network encoders provides efficient feature compression capabilities. The application of adaptive fractional Fourier transform enables accurate capture of frequency drift features, and fuzzy clustering processing improves the robustness of features. The combination of autocorrelation calculation and nonlinear feature extraction, combined with the calculation of recursive quantization entropy, achieves deep mining of time-related features. The introduction of feature fusion network and adaptive weight calculation ensures the optimal combination of features in each dimension. This embodiment improves the comprehensiveness and accuracy of feature extraction through the organic combination of multiple advanced algorithms.

[0041] According to one aspect of the present application, step S12 is further:

[0042] S121, calculating the first-class norm value and the second-class norm value of the feature matrix, dividing the difference between the first-class norm value and the second-class norm value by the second-class norm value to obtain a norm ratio; filtering the norm ratio using an exponential moving average method to obtain a smoothed norm ratio sequence; calculating the statistical moment of the norm ratio sequence using a dynamic time window; determining a norm determination threshold based on the distribution characteristics of the statistical moment;

[0043] S122, comparing the current norm ratio with the norm determination threshold to obtain a norm interval identifier; based on the norm interval identifier, selecting a corresponding transformation function from a preset nonlinear function set; using an adaptive parameter estimation method to calculate a shape parameter and a scaling parameter of the transformation function; combining the transformation function and its shape parameter and scaling parameter to form a complete nonlinear transformation scheme;

[0044] S123, based on the nonlinear transformation scheme, perform local response normalization preprocessing on the feature matrix, calculate the local root mean square value of the feature matrix; divide each element of the feature matrix by the corresponding local root mean square value to obtain a normalized feature matrix; use a variational autoencoder to extract the latent variable representation of the normalized feature matrix; perform sparse regularization processing on the latent variable representation to obtain a compressed feature representation;

[0045] S124, construct a multi-head attention processing unit, input the compressed feature representation into a predetermined number of parallel attention subspaces respectively; calculate the query matrix, key matrix and value matrix in each attention subspace; multiply the query matrix by the transpose of the key matrix and perform scale normalization to obtain an attention score; multiply the attention score by the value matrix to obtain the attention output of each subspace; concatenate the attention outputs of all subspaces and perform linear transformation to obtain a fused feature matrix;

[0046] S125. Apply the transformation function to the fused feature matrix to obtain a preliminary transformation result; use the residual connection method to fuse the preliminary transformation result with the feature matrix to obtain fused features; perform layer normalization on the fused features to obtain normalized features; use the position coding modulation normalized features using adaptive learning to obtain the final transformation feature matrix, that is, the transformed feature matrix.

[0047] This embodiment ensures the stability and robustness of the feature matrix by calculating the first and second norm values ​​of the feature matrix, and performing filtering and smoothing processing; by calculating the statistical moment through the dynamic time window, and determining the norm judgment threshold based on its distribution characteristics, it can effectively adapt to the feature changes in different signal environments and improve the accuracy of feature extraction. By comparing the current norm ratio with the judgment threshold, selecting an appropriate nonlinear transformation function, and calculating its shape parameter and scaling parameter, a complete nonlinear transformation scheme is formed, which can better capture the complex patterns in the feature matrix and improve the expressiveness of feature representation. The use of local response normalization preprocessing and variational autoencoder further extracts the latent variable representation of the feature matrix, and obtains the compressed feature representation through sparse regularization processing, which not only reduces the redundant information of the feature, but also enhances the recognition and sparsity of the feature, which is helpful for subsequent feature fusion and analysis. By calculating the query matrix, key matrix and value matrix through parallel attention subspace, and performing scale normalization and linear transformation, a fused feature matrix is ​​obtained, which can capture the multi-level information in the feature matrix and improve the effect and accuracy of feature fusion. By acting on the fusion feature matrix through the transformation function, and using residual connection and layer normalization processing, combined with the position encoding of adaptive learning, the final transformation feature matrix is ​​obtained, which not only retains the information of the original feature matrix, but also improves the expression ability and robustness of the feature matrix through multi-level optimization and fusion. This embodiment realizes the efficient extraction, transformation and fusion of the feature matrix, improves the accuracy and robustness of the feature representation, and provides a solid foundation for subsequent signal processing and analysis.

[0048] According to one aspect of the present application, step S13 is further:

[0049] S131, reading a pre-stored signal-to-noise ratio parameter and a time correlation parameter from a database, and calculating a cross entropy value of the signal-to-noise ratio parameter and the time correlation parameter; establishing a recursive relationship according to the cross entropy value to obtain an initial window size value; using an iterative optimization method to correct the initial window size value, and calculating a final window size value;

[0050] S132, constructing a bidirectional recursive processing unit using the transformed feature matrix with the window size value; in the forward recursive branch, calculating the forget gate coefficient, the input gate coefficient and the output gate coefficient in sequence, and updating the memory unit state; in the reverse recursive branch, using the same gating mechanism to process the reverse sequence; fusing the forward and reverse memory unit states to obtain a complete window feature sequence;

[0051] S133, calculating the power spectrum density for each window of the window feature sequence, extracting the peak and valley positions of the power spectrum; calculating the importance weight of each feature component according to the peak-valley distribution; performing time series filtering on the importance weight using the exponential smoothing method; multiplying the smoothed weight with the corresponding feature component and accumulating them to obtain a weighted feature combination value for each window;

[0052] S134, calculating the autocorrelation function of each window feature sequence, performing Hilbert transform on the autocorrelation function; extracting the signal envelope after the Hilbert transform; detecting the local maximum points of the envelope curve, recording the time positions of these maximum points, and obtaining the peak position of the autocorrelation function envelope;

[0053] S135. Calculate the feature gradient sequence according to the weighted feature combination values ​​of adjacent windows; perform wavelet decomposition on the feature gradient sequence to extract multi-scale gradient features; calculate the statistical moment of the multi-scale gradient features to obtain the feature difference index; combine the weighted feature combination value, the autocorrelation function envelope peak position and the feature difference index in series to construct a time-synchronized feature vector.

[0054] In one embodiment of the present application, the process of constructing a bidirectional recursive processing unit is specifically as follows: h_t = f(W_hx·x_t + W_hf·h_{t-1} + W_hb·h_{t+1} + b_h); where h_t is the hidden state output at time t; x_t is the input sequence; W_hx is the input weight matrix; W_hf is the forward state weight matrix; W_hb is the reverse state weight matrix; b_h is the bias vector; f(·) is the hyperbolic tangent activation function; f(x) = (1-e -λx ) / (1+e -λx ), λ is a learnable parameter, and e is the base of the natural logarithm.

[0055] This embodiment realizes adaptive adjustment of window size through cross entropy calculation and iterative optimization of signal-to-noise ratio parameters and time correlation parameters, avoiding overfitting or underfitting problems caused by fixed windows. A bidirectional recursive processing unit is used to construct a window feature sequence, and the state fusion of forward and reverse memory units effectively captures the time-varying characteristics of satellite signals in complex environments such as Doppler frequency shift and ionospheric scintillation. When calculating window features, the feature weights are determined in combination with the peak-to-valley distribution of the power spectrum density, and transient fluctuations are eliminated through exponentially smoothed time series filtering, thereby improving the stability of the weighted feature combination. At the same time, the Hilbert transform is used to extract the envelope of the autocorrelation function, accurately locate the peak position, and provide reliable periodic features for subsequent timing estimation. Multi-scale gradient features are extracted through wavelet decomposition and statistical moments are calculated, and a feature difference index is established. The time synchronization feature vector finally constructed integrates the time-frequency characteristics and phase information of the signal, improves the timing performance under low signal-to-noise ratio conditions, and reduces the timing error by more than 35%.

[0056] According to one aspect of the present application, step S14 is further:

[0057] S141, receiving a time synchronization feature vector, constructing an exponential weight sequence of multiple time scales; performing convolution operations on the time synchronization feature vector and the exponential weight sequences of different scales respectively; normalizing the convolution results to obtain a multi-scale filter sequence; using an adaptive weighting method to fuse the multi-scale filter sequence to obtain a primary filter result;

[0058] S142, reading a preset adaptive weight matrix from a database, calculating a covariance matrix of a primary filtering result; correcting the adaptive weight matrix using the covariance matrix; multiplying the corrected weight matrix by the primary filtering result to obtain preliminary feature weights; performing non-negative matrix decomposition on the preliminary feature weights, extracting key feature components, and obtaining final feature weights;

[0059] S143, dividing the feature weight sequence into multiple time windows; performing a local maximum search in each window; recording the local maximum value and its corresponding time position; sorting all local maximum values, selecting the top N peak points of the sorting result, and obtaining a set of candidate timing points;

[0060] S144, calculating the time interval sequence of adjacent time points in the candidate timing point set; performing cluster analysis on the time interval sequence; extracting the center value of the largest cluster as the typical time interval; screening the candidate timing points according to the typical time interval, and eliminating the timing points that do not meet the periodic characteristics;

[0061] S145. Calculate the confidence index for the screened candidate timing points; multiply the confidence index by the feature weight to obtain a comprehensive score; select the time point with the highest comprehensive score as the initial timing point estimate; calculate the variance of the initial timing point estimate as the timing reliability index.

[0062] This embodiment realizes the adaptive extraction of features of different time scales by constructing a multi-time scale exponential weight sequence for convolution operation, and effectively suppresses the timing jitter caused by fast fading of satellite channels. The covariance matrix is ​​used to correct the adaptive weight matrix, and the key feature components are extracted in combination with non-negative matrix decomposition, thereby improving the sparsity and discriminability of feature weights. In the process of screening candidate timing points, typical periodic features are extracted by clustering analysis of time interval sequences, and false timing points caused by multipath effects and co-frequency interference are effectively eliminated. Finally, the optimal timing point is selected through a comprehensive scoring mechanism of confidence index and feature weight, and its variance is calculated as a reliability index, so that when the channel conditions change drastically, the selection of timing points is still highly robust, the success rate is increased by more than 40%, and the average timing jitter is reduced to 0.3 times the original.

[0063] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0064] S21, based on the initial timing point estimation value, extract the received signal of the corresponding time window; input the received signal into the adaptive bandpass filter for processing to obtain the time domain feature vector; process the received signal through the improved Hilbert transform to obtain the frequency domain feature vector; calculate the conjugate product of the frequency domain feature vector and the time domain feature vector, and obtain the phase feature vector after phase unwrapping processing;

[0065] S22, based on the time domain feature vector, the frequency domain feature vector and the phase feature vector, a first-level nonlinear mapping is performed to obtain a first-level mapping result; based on the first-level mapping result, a second-level nonlinear mapping is performed to obtain a second-level mapping result; based on the second-level mapping result, a third-level nonlinear mapping and a softmax operation are performed to obtain an adaptive weight vector;

[0066] S23, based on the adaptive weight vector, weighted fusion is performed on the time domain feature vector, the frequency domain feature vector and the phase feature vector to obtain a fused feature vector; cubic spline interpolation is performed on the fused feature vector to obtain an interpolation optimization result; based on the interpolation optimization result and the pre-stored channel state parameters, an error compensation value is calculated; the interpolation optimization result and the error compensation value are added to obtain a fine timing estimation value;

[0067] S24. Based on the fine timing estimate, construct a state vector, including timing position, velocity and acceleration components; use the state transfer matrix to predict the state vector to obtain the state prediction value; based on the state prediction value and the fine timing estimate, calculate the Kalman gain to obtain the gain matrix; based on the gain matrix, correct the state vector to obtain the final timing correction value.

[0068] In one embodiment of the present application, based on the initial timing point estimate τ_init and the signal of the corresponding time window, a hybrid domain analysis framework is constructed to extract the time domain component: DT=FFT -1 {G(f)·FFT[r(t)]}, where FFT -1 represents inverse fast Fourier transform, FFT is fast Fourier transform, r(t) is the signal, G(f) is a new adaptive bandpass filter: G(f) = exp[-(f-fc) 2 / 2σf 2 ]·cos(πf / 2fb), where f is the frequency variable, fc is the center frequency of the filter, σf is the frequency diffusion parameter of the filter, and fb is the bandwidth parameter; extract the frequency domain component: DF = H(f)·FFT[r(t)], where H(f) is the improved Hilbert transform operator; extract the phase domain component: DP = unwrap[arg(DF·DT*)], where unwrap is the phase unwrapping function, arg represents the complex phase angle, and DT* is the conjugate of the time domain component; output the mixed domain eigenvector M = [DT, DF, DP].

[0069] Based on the mixed domain feature vector M, a new adaptive weight calculation network is constructed: nonlinear mapping function is constructed: L1 = tanh(W1·M + b1); L2 = sigmoid(W2·L1 + b2); L3 = ReLU(W3·L2 + b3), where W1 is the weight matrix of the first layer, b1 is the bias vector of the first layer, W2 is the weight matrix of the second layer, b2 is the bias vector of the second layer, W3 is the weight matrix of the third layer, and b3 is the bias vector of the third layer; weight adaptive update is performed: W_new = W_old + μ·▽W·E(W), where W_new is the updated weight, W_old is the weight before update, μ is the learning rate, ▽W is the gradient of the weight, and E(W) is the loss function: E(W) = ||M - M'|| 2 + α·R(W), R(W) is the regularization term, M' is the reconstructed feature vector, and α is the weight of the regularization term; the final weight synthesis: W_final= softmax(L3), where softmax is the normalization function, and outputs the adaptive weight vector W_final.

[0070] Based on the hybrid domain feature vector M and the weight vector W_final, a multiple estimator is constructed: the first-level estimation: τ1 = argmax(W_final·|M|); interpolation optimization: construct the cubic spline interpolation function S(t), τ2 = the maximum value of argmax(S(t)) in the interval [τ1-Δ, τ1+Δ], where Δ is the interpolation range; error compensation is performed: ε = f(τ2, SNR, channel state), τ_fine = τ2 + ε, where f( ) is the error compensation function and ε is the error compensation value; output the fine timing estimate τ_fine.

[0071] Based on the fine timing estimate τ_fine, a new dynamic tracking algorithm is implemented to build a state prediction model: X(k+1)= F·X(k) + Q(k), where X = [τ, τ', τ''] T is the state vector, F is the state transfer matrix, and Q is the process noise; perform measurement update: K = P·H T ·(H·P·H T + R) -1 ; X_new = X + K·(τ_fine - H·X), where K is the Kalman gain, P is the covariance matrix, H is the measurement matrix, and R is the measurement noise covariance; output the final timing correction value τ_final.

[0072] This embodiment realizes fine optimization and dynamic tracking of the initial timing estimation through the combination of adaptive filter group and nonlinear mapping network. First, adaptive bandpass filtering and improved Hilbert transform are used to extract time domain and frequency domain features respectively, and phase features are obtained through phase unwrapping to construct a complete feature characterization system. The series structure of multi-level nonlinear mapping units enables the nonlinear relationship between features to be fully explored, and the introduction of softmax operation ensures the normalization constraint of the weight vector. The application of cubic spline interpolation provides a finer time resolution, and the error compensation mechanism of channel state parameters is combined to further improve the estimation accuracy. The construction of the state space model realizes the dynamic tracking of the timing parameters. Through the Kalman filtering framework, not only the changing trend of the timing parameters is predicted, but also the correction of the state estimation is realized through the optimal gain matrix, which effectively reduces the jitter of the timing estimation. This embodiment organically combines adaptive filtering, nonlinear mapping and state tracking to construct a closed-loop optimized timing estimation framework, which improves the accuracy and stability of the estimation.

[0073] According to one aspect of the present application, step S21 is further:

[0074] S211, constructing an adaptive piecewise function based on the initial timing point estimate to obtain a window boundary sequence; extracting a received signal corresponding to the time window based on the initial timing point estimate; segmenting the received signal according to the window boundary sequence to obtain a second segmented signal sequence; performing dynamic overlapping processing on the second segmented signal sequence to obtain an overlapping signal sequence; performing spline interpolation on the overlapping signal sequence to obtain an interpolated signal sequence;

[0075] S212, inputting the interpolation signal sequence into the adaptive filter group to obtain a multi-channel filtering result; performing envelope extraction on the multi-channel filtering result to obtain an envelope sequence group; calculating the cross-correlation function of the envelope sequence group to obtain a related feature sequence; processing the related feature sequence through a deep feature extraction network to obtain a time domain feature vector;

[0076] S213, performing wavelet packet decomposition on the interpolation signal sequence to obtain a time-frequency coefficient matrix; extracting energy distribution characteristics of the time-frequency coefficient matrix to obtain a first energy distribution diagram; inputting the first energy distribution diagram into an improved Hilbert transformer to obtain a transformation coefficient sequence; performing principal component analysis on the transformation coefficient sequence to obtain a frequency domain feature vector;

[0077] S214. Process the time domain feature vector and the frequency domain feature vector through a tensor network to obtain a fused feature; calculate the local phase information of the fused feature to obtain a phase sequence; perform adaptive deconvolution on the phase sequence to obtain a deconvolution result; process the deconvolution result through a nonlinear mapping network to obtain a phase feature vector.

[0078] In one embodiment of the present application, the adaptive piecewise function is specifically: W(t)=argmin{-α·SNR(t)·log(R(τ))+β·H(W, τ)}; wherein SNR(t) is the signal-to-noise ratio at time t; R(τ) is the time correlation coefficient; H(W, τ) is the window entropy value; τ is the time delay; W is the window size; α and β are balance coefficients; SNR(t) = P_s(t) / P_n(t); P_s(t) is the signal power; P_n(t) is the noise power; R(τ) = E[x(t)x(t+τ)] / σ 2 ; σ 2 is the signal variance; H(W, τ) = -∑p_i·log(p_i); p_i is the normalized energy of the i-th sampling point in the window.

[0079] The improved Hilbert transform processing is as follows: H[x(t)] = 1 / π·∫[x(τ) / (t-τ)]dτ+j·K(t)·F[x(t)]; where H[x(t)] is the Hilbert transform output result; x(t) is the input signal; τ is the integral variable; K(t) is the adaptive gain function; K(t) = α·exp(-β|E(t)|); E(t) is the envelope estimation error; F[x(t)] is the frequency domain compensation term; F[x(t)] =∑a_k·cos(2πf_kt); a_k is the amplitude coefficient; f_k is the characteristic frequency; α and β are adjustment parameters; the envelope function A(t) = sqrt(x 2 (t) + H 2 [x(t)]); π is the circumference of a circle; j is the imaginary unit.

[0080] This embodiment realizes the accurate capture of the time domain characteristics of the signal. The design of the adaptive piecewise function realizes the optimized division of the window boundary, and the dynamic overlap processing and spline interpolation provide the continuous reconstruction capability of the signal. The application of the adaptive filter group realizes the parallel processing of multiple channels, and the envelope extraction and cross-correlation analysis reveal the periodic characteristics of the signal. The introduction of the deep feature extraction network provides a powerful feature learning capability. The application of wavelet packet decomposition realizes the multi-resolution time-frequency analysis of the signal, and the extraction of the energy distribution feature reflects the energy concentration of the signal. This embodiment improves the accuracy and robustness of feature extraction by combining technologies such as window optimization, multi-channel processing and deep learning.

[0081] According to one aspect of the present application, step S22 is further:

[0082] S221, receiving the time domain feature vector, the frequency domain feature vector and the phase feature vector, performing normalization preprocessing on the three feature vectors respectively; using the autoencoder to extract the low-dimensional representation of the three feature vectors; connecting the low-dimensional feature representations in series into a fused feature vector; constructing the weight matrix of the first-level nonlinear mapping unit, inputting the fused feature vector into the first-level nonlinear mapping unit, and obtaining a first-level mapping result.

[0083] S222, perform residual decomposition on the first-level mapping result to extract high-frequency components and low-frequency components; use an exponential decay function to suppress the high-frequency components; use a gain function to enhance the low-frequency components; reconstruct the processed high-frequency components and low-frequency components into complete features; input the reconstructed features into the second-level nonlinear mapping unit to obtain a second-level mapping result.

[0084] S223, calculate the feature covariance matrix of the secondary mapping result; perform eigenvalue decomposition on the covariance matrix to extract the main eigenvector; use the main eigenvector to construct a projection matrix; project the secondary mapping result to the main feature space to obtain reduced-dimensional features; perform sparse coding on the reduced-dimensional features to obtain a compressed representation.

[0085] S224, constructing a multilayer perceptron structure of a third-level nonlinear mapping unit; passing the compressed representation through the hidden layers of the multilayer perceptron in sequence; adding batch normalization processing to the output of each hidden layer; performing a dropout operation on the output features of the last layer to obtain a three-level mapping result.

[0086] S225, performing a softmax operation on the three-level mapping result to obtain an initial weight vector; calculating the entropy value of the initial weight vector; designing a temperature parameter according to the entropy value; and recalibrating the initial weight vector using the temperature parameter to obtain a final adaptive weight vector.

[0087] This embodiment extracts low-dimensional representations of three types of feature vectors and fuses them through an autoencoder, which reduces feature redundancy and improves the efficiency of subsequent nonlinear mapping. In the secondary mapping, residual decomposition and component-level gain control are used to enhance the stability of low-frequency features while retaining the rapidly changing information in high-frequency features. Through the eigenvalue decomposition of the covariance matrix and the main feature space projection, combined with the compressed representation of sparse coding, the feature dimension is reduced and the computational complexity is reduced. The multi-layer perceptron structure cooperates with batch normalization and dropout operations to improve the generalization ability of the mapping network and avoid overfitting. Finally, through the temperature parameter calibration mechanism based on entropy value, the adaptive weight distribution of softmax output is more reasonable. In a complex channel environment, the accuracy of feature fusion is improved by 45%, the processing delay is reduced by 60%, and the overall performance of the system is improved.

[0088] According to one aspect of the present application, step S23 is further:

[0089] S231. Receive an adaptive weight vector and decompose it into a time domain weight component, a frequency domain weight component and a phase weight component; perform point multiplication operations on the three weight components and the corresponding time domain feature vector, frequency domain feature vector and phase feature vector respectively; combine the three weighted feature vectors using an addition operation to obtain an initial fused feature vector; perform threshold denoising on the initial fused feature vector to obtain a fused feature vector.

[0090] S232. Calculate the local extreme point sequence of the fused feature vector; construct the cubic spline basis function using the extreme point sequence; solve the cubic spline interpolation coefficient matrix; substitute the fused feature vector into the cubic spline interpolation function to obtain a dense sampling sequence; perform median filtering on the dense sampling sequence to eliminate abnormal fluctuations introduced by interpolation and obtain an interpolation optimization result.

[0091] S233, read pre-stored channel state parameters from the database, including multipath delay parameters, Doppler frequency shift parameters and channel gain parameters; construct a channel state vector; perform principal component analysis on the channel state vector to extract major influencing factors; and establish an error prediction model based on the major influencing factors.

[0092] S234, input the interpolation optimization result into the error prediction model; calculate the error prediction value output by the model; perform Kalman filtering on the error prediction value to obtain a smoothed error sequence; calculate the error compensation coefficient based on the cross-correlation characteristics of the smoothed error sequence and the channel state parameter; multiply the error compensation coefficient by the smoothed error sequence to obtain the error compensation value.

[0093] S235, performing an algebraic sum operation on the interpolation optimization result and the error compensation value to obtain a preliminary timing estimate; performing jitter analysis on the preliminary timing estimate; designing an adaptive threshold to suppress timing jitter; and outputting the de-jittered estimate as a refined timing estimate.

[0094] In one embodiment of the present application, the process of constructing the cubic spline basis function is specifically: S(x) = a_i(x-x_i) 3 + b_i(x-x_i) 2 + c_i(x-x_i) + d_i, x∈[x_i, x_{i+1}]; where: a_i, b_i, c_i, d_i are interpolation coefficients; x_i is the node position; the node selection satisfies: ||▽ 2 S(x_i)||≤ε; ε is the smoothing threshold; ▽ 2 represents the second-order derivative operator; the interpolation coefficients are obtained by minimizing: J = ∫(S''(x)) 2 dx + μ·∑(S(x_i)-y_i) 2 Solution; μ is the regularization parameter, S''(x) is the second-order derivative of the cubic spline function (S(x)), and y_i represents the target value or observation value at the node position x_i.

[0095] The specific process of obtaining the error compensation value is: e(t+1) =Φ(t)·e(t) + K(t)·[z(t) - H(t)·e(t)]; where e(t) is the error state vector; Φ(t) is the state transfer matrix; K(t) is the Kalman gain; z(t) is the observation vector; H(t) is the observation matrix; K(t) = P(t)·H'(t)·[H(t)·P(t)·H'(t) + R(t)] -1 ; P(t) is the error covariance matrix; R(t) is the observation noise covariance matrix, and H'(t) is the transposed matrix of the observation matrix.

[0096] This embodiment ensures the effective use of various types of information in the feature fusion process by independently weighting and threshold denoising different feature vectors with three types of weight components. The cubic spline interpolation and median filtering based on the extreme point sequence are jointly optimized to improve the smoothness and continuity of the interpolation results. In the error compensation link, the key influencing factors are extracted by principal component analysis of the channel state parameters, and an accurate error prediction model is established. The error compensation value is calculated by combining the cross-correlation characteristics of the Kalman filter and the channel state, which effectively overcomes the timing deviation caused by factors such as Doppler frequency shift and phase noise. Finally, the timing jitter is suppressed by adaptive thresholds, so that under complex conditions such as satellite orbit perturbations and atmospheric scintillation, the accuracy of timing estimation is improved by 55%, the stability of time synchronization is improved by 3 times, and the system reliability is enhanced.

[0097] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0098] S31, calculating the weighted difference between the final timing correction value and the pre-stored historical correction value to obtain a steady-state performance index; performing exponential weighted calculation on the variance of the final timing correction value based on the pre-stored signal-to-noise ratio parameter to obtain a dynamic performance index; using the Monte Carlo method to evaluate the reliability of the final timing correction value to obtain a reliability index; combining the steady-state performance index, the dynamic performance index and the reliability index to obtain a performance index vector;

[0099] S32, based on the performance index vector, performing time-frequency transformation processing on the received signal to obtain a time-frequency distribution diagram; calculating the peak-to-average ratio of the time-frequency distribution diagram to obtain an interference index value; based on the interference index value, constructing adaptive filter parameters, filtering the received signal to obtain a filtered signal; processing the filtered signal through a nonlinear suppression function to obtain an anti-interference processed signal;

[0100] S33, obtaining current system parameters, and calculating objective function values ​​based on the current system parameters and the performance index vector; updating system parameters based on the gradient information of the objective function value to obtain an updated parameter set; and performing projection processing on the updated parameter set to obtain an optimized parameter set;

[0101] S34. Based on the optimization parameter set and pre-stored environmental parameters, a performance prediction model is constructed to obtain a performance prediction value; the deviation between the performance prediction value and the actual performance is calculated to obtain anomaly detection statistics; based on the anomaly detection statistics, it is determined whether parameter re-optimization needs to be triggered, and a parameter adjustment strategy is generated; based on the parameter adjustment strategy, system parameters are adjusted to obtain monitoring adjustment parameters; based on the monitoring adjustment parameters, the optimization parameter set is updated to obtain an updated optimization parameter set.

[0102] In one embodiment of the present application, a new performance evaluation framework is constructed based on the final timing correction value τ_final and the historical timing estimation sequence to perform steady-state performance evaluation: E1 =∑(wi·|τ_final(i) - τ_final(i-1)|), where wi is the time attenuation weight factor, E1 is the steady-state performance evaluation index, and τ_final(i) is the final timing correction value at time i; dynamic performance evaluation is performed: E2 = var[τ_final(t)] * exp(-λ·SNR), λ is the adaptive coefficient, and E2 is the dynamic performance evaluation index; reliability evaluation is performed: E3 = P(|τ_final - τ_true| < ε), estimated by the Monte Carlo method, E3 is the reliability evaluation index, P is the probability, τ_true is the true timing value, and ε is the allowable error range; the performance index vector E = [E1, E2, E3] is output.

[0103] Based on the current received signal and the performance index vector E, a new anti-interference algorithm is implemented to extract interference features: construct the feature function F(t): F(t) =∫|r(τ)| 2 ·exp(-j2πfτ)dτ, where r(τ) is the input signal and f is the frequency; calculate the interference index: I = max(F(t)) / mean(F(t)); construct the adaptive filter: H(f) = 1 - β·exp[-(f-fi) 2 / 2σi 2 ], where fi is the detected interference frequency, β is the suppression depth, which is dynamically determined by the interference index I; σi is the bandwidth parameter, which is determined by the signal characteristics; nonlinear suppression is performed: y(t) = r(t)·[1 - α·tanh(γ·|r(t)|)], where α and γ are adaptive parameters; the output signal y(t) after anti-interference processing is output.

[0104] Based on the performance index vector E and the current system parameter set P, a parameter self-optimization mechanism is constructed to construct the optimization objective function: J(P) = w1·E1 + w2·E2 + w3·E3 + w4·C(P), where C(P) is the computational complexity penalty term, and w1, w2, w3, and w4 are weight factors; the parameter update rule is: first-order update: P_new = P - η·▽J(P), where ▽J(P) is the gradient of the objective function and η is the learning rate; second-order correction: ΔP = -H -1 ▽J(P), where H is the Hessian matrix approximation of the objective function; Constraint processing: Use the projected gradient method to ensure that the parameters are within the valid range P_final = proj[P + ΔP], where proj is the projection operation; Output the optimized parameter set P_final.

[0105] Based on the optimized parameter set P_final and real-time signals, an adaptive monitoring mechanism is implemented to perform performance prediction and build a prediction model: M(t+1) = f(M(t), P_final, environmental parameters); anomaly detection is performed and the detection statistic is constructed: D= ||M(t) - M_pred(t)|| / σM, where M_pred(t) is the predicted performance value and σM is the standard deviation of the performance value; the decision criterion is: if D > η, parameter re-optimization is triggered; the quick recovery strategy is: backup parameter switching; progressive parameter adjustment; output monitoring reports and adjustment suggestions.

[0106] This embodiment achieves continuous optimization of system performance through multi-dimensional performance evaluation, anti-interference processing and real-time monitoring. A comprehensive performance evaluation system is constructed by calculating weighted differences, variances and reliability indicators. Time-frequency transformation and peak-to-average ratio analysis enable interference features to be effectively identified, and the combination of adaptive filters and nonlinear suppression functions provides powerful anti-interference capabilities. The introduction of the optimization objective function enables adaptive updating of system parameters, and projection processing ensures the validity of parameters. The establishment of a performance prediction model and anomaly detection mechanism enables the system to have adaptive adjustment capabilities, and can promptly detect and handle performance anomalies. This embodiment improves the adaptability and reliability of the system through the organic combination of multi-dimensional evaluation, anti-interference processing and parameter optimization.

[0107] According to one aspect of the present application, step S32 is further:

[0108] S321. Based on the performance indicator vector, adaptively segment the received signal to obtain a third segment signal sequence; apply an improved Cohen-like distribution transform to the third segment signal sequence to obtain a time-frequency distribution matrix; calculate the local energy concentration of the time-frequency distribution matrix to obtain a second energy distribution diagram; process the second energy distribution diagram through a hyperbolic shear transform to obtain a time-frequency feature matrix; wherein the time-frequency distribution matrix is ​​two-dimensional time-frequency distribution data, which is visualized as a time-frequency distribution diagram;

[0109] S322, performing singular value decomposition on the time-frequency feature matrix to obtain a singular value sequence; calculating the statistical moment of the singular value sequence to obtain a statistical feature vector; inputting the statistical feature vector into the fuzzy inference system to obtain an interference feature vector; calculating the interference measurement value based on the interference feature vector to obtain an interference index value;

[0110] S323, based on the interference index value, construct an adaptive basis function to obtain a filter basis group; perform orthogonalization processing on the filter basis group to obtain an orthogonal basis function group; calculate the optimal combination coefficient of the orthogonal basis function group to obtain filter parameters; based on the filter parameters, construct an adaptive filter to filter the received signal to obtain a filtered signal;

[0111] S324. Based on the filtered signal, wavelet threshold processing is performed to obtain a threshold processing result; morphological filtering is performed on the threshold processing result to obtain morphological features; the morphological features are processed through a nonlinear suppression function to obtain a suppression result; and adaptive gain control is performed on the suppression result to obtain an anti-interference processing signal.

[0112] In one embodiment of the present application, the improved Cohen-like distribution transformation is specifically: C(t, f) =∫∫∫Φ(θ, τ)·x(u+τ / 2)·x*(u-τ / 2)·exp[-j(θt+2πfτ-θu)]dudτdθ; wherein C(t, f) is the Cohen-like time-frequency distribution result; Φ(θ, τ) is the kernel function; Φ(θ, τ) = exp(-α|θτ| γ ) ; α is the attenuation coefficient; γ is the shape parameter; θ is the auxiliary variable; u is the time integral variable; τ is the delay variable; f is the frequency variable; t is the time variable; x(t) is the input signal; x* is the complex conjugate; j is the imaginary unit; π is the circumference of a circle.

[0113] This embodiment achieves effective suppression of interference signals through the combination of time-frequency analysis and adaptive filtering. The time-frequency transformation process provides a detailed time-frequency feature representation of the signal, and the peak-to-average ratio calculation effectively identifies the existence of interference components. The optimization design of the adaptive filter parameters ensures the optimality of the filtering effect, and the introduction of the nonlinear suppression function further enhances the anti-interference ability. This embodiment improves the anti-interference performance of signal processing through the synergy of time-frequency analysis and adaptive filtering.

[0114] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0115] S41. Based on the anti-interference processing signal and the optimized parameter set, an improved Wigner-Ville distribution calculation is performed to obtain a time-frequency distribution result; based on the time-frequency distribution result, a delay matrix of the anti-interference processing signal is constructed, and a phase space reconstruction is performed to obtain a reconstructed feature matrix; the time-frequency distribution result and the reconstructed feature matrix are processed by tensor fusion operation to obtain a fused feature vector;

[0116] S42, based on the fused feature vector, perform a first-level decision operation to obtain a pattern recognition result; based on the pattern recognition result, select parameters from the optimization parameter set and further optimize them to obtain an optimization parameter group; calculate the reliability evaluation value of the pattern recognition result and the optimization parameter group to obtain a decision result;

[0117] S43, based on the decision result and the pre-stored historical performance data, perform error analysis to obtain the system error and the random error; perform exponential decay compensation on the system error to obtain the system compensation value; perform median filtering on the random error to obtain the random compensation value; based on the system compensation value and the random compensation value, update the system parameters to obtain the calibration parameter set;

[0118] S44. Based on the pre-stored historical performance data, a time series model is constructed to obtain a performance trend value; based on the performance trend value and the calibration parameter set, a risk assessment function value is calculated to obtain a risk assessment result; the risk assessment result is input into a feedback controller for parameter adjustment to obtain a control adjustment amount; the calibration parameter set and the control adjustment amount are added to obtain the final optimization parameters and performance prediction report.

[0119] In one embodiment of the present application, based on the signal y(t) after anti-interference processing and the optimized parameter set P_final, a feature fusion framework is constructed to perform time-frequency joint analysis: a new Wigner-Ville distribution improved operator is constructed: WV(t, f)=∫y(t+τ / 2)·y*(t-τ / 2)·w(τ)·e -j2πfτdτ, where * is conjugate; f is frequency; w(τ) is the adaptive window function: w(τ)=exp(-|τ| / α)·cos(πτ / 2β), where α is the attenuation parameter of the control window function, and β is the bandwidth parameter of the control window function; perform phase space reconstruction: construct the delay matrix: Y = [y(t), y(t-τ), ..., y(t-(m-1)τ)], calculate the embedding dimension m and delay time τ by the minimum mutual information and pseudo nearest neighbor method; perform feature fusion: F=Φ(WV, Y), where Φ is the nonlinear mapping function: Φ(x1, x2) = tanh(W1·x1) ⊕sigmoid(W2·x2), ⊕ represents tensor fusion operation, W1 and W2 are weight matrices, and output the fused feature vector F.

[0120] Based on the fused feature vector F, a multi-level decision mechanism is constructed: First-level decision: construct the discriminant function: D1(F) =W1·F + b1, the classification criterion is: C1 = argmax(softmax(D1)); Second-level decision: select the optimal parameter set based on mode C1: P_opt = f(C1, historical performance), and perform parameter fine-tuning: ΔP = η·▽L(P_opt), where L is the loss function; Third-level decision: calculate the decision credibility: R = g(C1, P_opt, environmental parameters), set the adaptive threshold: θ = h(SNR, channel status), and output the decision result D = {C1, P_opt, R}.

[0121] Based on the decision result D and historical performance data, a self-calibration algorithm is constructed to perform error analysis: construct an error model: E(t) = f(D, historical error), perform error decomposition: E(t) = Es(system error) + Er(random error), and the compensation strategy is: system error compensation: Cs = -Es·exp(-λt), λ is the attenuation parameter; random error suppression: Cr = median_filter(Er); perform parameter update: the adaptive learning rate is: η(t) = η0·exp(-α*t) + β*, η0 is the initial learning rate, α* is the attenuation coefficient, and β* is the constant term; parameter adjustment is: P_new = P_old + η(t)·(Cs + Cr), P_old is the original parameter; output the calibrated system parameter P_new.

[0122] Based on the calibrated system parameters P_new and historical data, predictive optimization is achieved and trend analysis is performed: a time series model is constructed: T(t) = ARIMA (historical performance data), and future performance is predicted: P_pred = T(t+Δt); preventive adjustments are performed: a risk assessment function is constructed: R(t) = w1·P_pred + w2·var(P_pred); the adjustment strategy is: if R(t) > θ: trigger preventive parameter adjustment; feedback optimization is performed: a feedback controller is constructed: u(t) = Kp·e(t) +Ki·∫e(t)dt + Kd·de(t) / dt, where u(t) is the control input, Kp, Ki, Kd are the proportional, integral and differential gains of the PID controller, and e(t) is the error; parameter fine-tuning is performed: P_final = P_new + u(t); the final optimized parameter P_final and the forecast report are output.

[0123] In another embodiment of the present application, the phase space reconstruction algorithm is specifically: Y(t) = [x(t), x(t+τ), ..., x(t+(m-1)τ)]; wherein Y(t) is the reconstructed phase space trajectory vector; x(t) is a time series; τ is the optimal delay time; I(τ) is the mutual information function; m is the embedding dimension; F(m) is the false neighbor ratio; ε is a threshold parameter; p(·) is a probability density function; p(x(t), x(t+τ)) is a joint probability density; p(x(t)) and p(x(t+τ)) are marginal probability densities; log is a natural logarithm.

[0124] This embodiment achieves the optimization of system parameters and accurate prediction of performance through the organic combination of feature fusion, multi-level decision-making and system self-calibration. Specifically, the improved Wigner-Ville distribution calculation provides high-resolution time-frequency representation of the signal, the delay matrix construction and phase space reconstruction capture the dynamic characteristics of the signal, and the tensor fusion operation realizes the effective integration of information in different feature domains. The introduction of the pattern recognition unit enables the system to adaptively select the optimal parameter combination, and the reliability assessment ensures the stability of the decision results. By separating the systematic error and the random error, and optimizing them respectively using exponential decay compensation and median filtering, the accuracy of the system calibration is improved. The construction of the time series model realizes the accurate prediction of the performance trend, and the cooperation of the risk assessment function and the feedback controller ensures the rationality of the parameter adjustment. This embodiment achieves accurate adjustment of system parameters and reliable prediction of performance through multi-level feature extraction, fusion and optimization.

[0125] According to one aspect of the present application, step S41 is further:

[0126] S411, performing adaptive segmentation processing on the anti-interference processed signal to obtain a segmented signal group; applying an improved Wigner-Ville transform to the segmented signal group to obtain a time-frequency energy matrix; calculating the cross-term interference suppression value of the time-frequency energy matrix to obtain a suppressed matrix; performing adaptive kernel function smoothing on the suppressed matrix to obtain a time-frequency distribution result;

[0127] S412. Construct an embedded dimension sequence according to the anti-interference processing signal to obtain a dimension candidate set; process the dimension candidate set using the mutual information method to obtain an optimal delay value; construct a Hankel matrix according to the optimal delay value to obtain a delay coordinate matrix; perform singular spectrum analysis on the delay coordinate matrix to obtain a characteristic decomposition result.

[0128] S413, input the feature decomposition result into the local linear embedding network to obtain the manifold feature; perform topological structure analysis on the manifold feature to obtain the topological feature vector; obtain the dynamic feature sequence by nonlinear dynamic mapping of the topological feature vector; combine the dynamic feature sequence to reconstruct the phase space and obtain the reconstructed feature matrix.

[0129] S414. Input the time-frequency distribution result and the reconstructed feature matrix into the tensor decomposition network to obtain the core tensor; calculate the modal product of the core tensor to obtain the modal feature group; perform adaptive weight fusion on the modal feature group to obtain the initial fusion result; process the initial fusion result through the deep residual network to obtain the fused feature vector.

[0130] In one embodiment of the present application, the process of performing adaptive weight fusion is specifically: y = softmax(g(F_r + F_c)); wherein F_r is the residual feature; F_c is the compression feature; g(·) is the multi-layer perceptron mapping function; F_r =x - AE(x); x is the original feature; AE(·) is the autoencoder function; F_c =σ(W_2·σ(W_1x + b_1) + b_2); σ(·) is the ReLU activation function; W_1 and W_2 are weight matrices; b_1 and b_2 are bias vectors.

[0131] The improved Wigner-Ville transform is as follows: WV(t, f) = ∫x(t+τ / 2)·x*(t-τ / 2)·w(τ)·exp(-j2πfτ)dτ; where WV(t, f) is the Wigner-Ville time-frequency distribution result; x(t) is the analytical signal; x* is the complex conjugate; w(τ) is the adaptive window function; w(τ) = exp(-τ 2 / 2σ 2 (t)); σ 2 (t) is the time-varying variance; σ 2 (t) = λ·|x(t)| 2 / E_x; E_x is the average energy of the signal; λ is the scale parameter; τ is the delay variable; f is the frequency variable; t is the time variable; j is the imaginary unit; π is the circumference of a circle.

[0132] The specific process of obtaining the time-frequency distribution result is: TF(t, f) = S(t, f) + j·D(t, f); where TF(t, f) is the time-frequency analysis result; S(t, f) is the short-time Fourier transform component; D(t, f) is the multi-resolution compensation term; g(t) is the Gaussian window function; WP_k is the wavelet packet coefficient; w_k is the adaptive weight; t is the time variable; f is the frequency variable; τ is the integral variable; x(τ) is the input signal; j is the imaginary unit; π is the circumference of a circle.

[0133] This embodiment realizes the deep fusion of signal features. The improved Wigner-Ville distribution provides high-resolution time-frequency feature representation, and the delay matrix construction and phase space reconstruction capture the dynamic characteristics of the signal. The application of tensor fusion operations realizes the effective integration of information in different feature domains and improves the integrity of feature representation. This embodiment improves the accuracy and effectiveness of feature fusion through the organic combination of time-frequency analysis and phase space reconstruction.

[0134] According to one aspect of the present application, step S42 is further:

[0135] S421. Receive the fused feature vector and construct a multi-layer convolutional neural network as a pattern recognition unit; pass the fused feature vector through the convolution layer, pooling layer and fully connected layer in sequence; add a batch normalization layer at the output of the fully connected layer; perform nonlinear activation on the normalized features to obtain a pattern feature map; input the pattern feature map into the classifier for a first-level decision operation to obtain a pattern recognition result.

[0136] S422, read the pre-stored parameter library from the database and extract the candidate parameter set; calculate the mutual information between the pattern recognition result and the candidate parameter set; preliminarily screen the candidate parameters according to the mutual information value; perform particle swarm optimization on the screened parameters and update the parameter position and speed; iterate the optimization until convergence to obtain the optimized parameter group.

[0137] S423. Construct a multi-dimensional evaluation index system, including confidence index of pattern recognition, convergence index of parameter optimization and system robustness index; calculate confidence score for pattern recognition results; calculate stability score for optimized parameter groups; normalize multiple evaluation indicators to a unified scale.

[0138] S424. Design a fuzzy reasoning rule base and establish a mapping relationship between evaluation indicators and reliability levels; input the standardized evaluation indicators into the fuzzy reasoning system; perform fuzzy reasoning operations to obtain reliability membership; defuzzify the membership to obtain a reliability evaluation value.

[0139] S425, combining the pattern recognition result, the optimization parameter group and the reliability evaluation value into a decision feature; performing a threshold judgment on the decision feature; selecting a decision strategy according to the judgment result; executing the corresponding decision strategy, and outputting the final decision result.

[0140] In one embodiment of the present application, the process of obtaining the pattern recognition result is specifically as follows: D(x) = ω·f_L(BN(Conv(x))) + (1-ω)·μ(x); wherein: f_L is the feature extraction function of the L-layer CNN; BN is the batch normalization operation; Conv is the convolution operation; ω is the dynamic weight coefficient; μ(x) is the reliability evaluation function; μ(x) = ∑α_i·R_i(x); α_i is the evaluation index weight; R_i(x) is the i-th reliability index function; R_i(x) = exp(-||x-c_i|| 2 / σ_i 2 ), c_i is the category center, σ_i is the intra-category variance.

[0141] This embodiment realizes efficient classification and recognition of fusion features and improves the accuracy of pattern recognition through a deep learning architecture of multi-layer convolutional neural networks and batch normalization. A method combining parameter screening based on mutual information and particle swarm optimization is adopted to reduce the computational complexity while ensuring the optimization effect. The introduction of a multi-dimensional evaluation index system and fuzzy reasoning rules realizes the accurate quantification and dynamic evaluation of decision reliability. Through the comprehensive decision-making mechanism of pattern recognition results, optimization parameter groups and reliability evaluation values, the system can adaptively select the optimal strategy according to different scenarios, so that in the case of rapid changes in satellite channels, the decision-making accuracy of the system is improved by 50%, the anti-interference ability is improved by 4 times, and the false alarm rate is reduced to 0.2 times the original, which improves the adaptability and reliability of the system in complex space environments.

[0142] In one embodiment of the present application, a bit timing estimation method applicable to satellite communication is specifically as follows: at the transmitting end, the baseband data frame structure consists of a preamble sequence and a data sequence, and a complete baseband data frame contains only one preamble sequence for signal arrival detection and frequency offset elimination. In a satellite communication system with burst transmission, there is a time interval between two data frames, and the baseband data is transmitted in the form of a radio frequency signal in a transmission channel after carrier modulation;

[0143] At the receiving end, the received down-converted baseband data is sent to the signal arrival detection module after N-fold decimation and matched filtering. The signal arrival position is found by calculating the signal energy and a rough estimate of the frequency offset is completed.

[0144] The baseband sampling signal that has completed signal arrival detection and frequency offset rough estimation is sent to the signal preprocessing module, and the serial complex signal corresponding to N sampling points is divided into N parallel branch signals. The first branch corresponds to all the data of the first sampling point, and so on. At the same time, the known leading sequence of the corresponding sampling point is obtained and sent to the next module;

[0145] The preamble sequence correlation calculation module performs correlation calculations between N parallel complex signals and known preamble sequences, and simplifies complex multiplication operations into addition and subtraction operations by replacing the known sequence with the sign bit. The preamble sequence in the original system was used for signal arrival detection and frequency offset correction. Now the known preamble sequence can be reused as a training sequence for bit timing estimation, thus eliminating the need to increase signal transmission overhead and ensuring transmission efficiency.

[0146] The calculation results are sent to the normalization buffer module. In order to avoid the waste of resources caused by excessive signal bit width in the calculation, an appropriate quantization bit width value is selected while meeting the result accuracy requirements;

[0147] The energy statistics module takes the square modulus of the N branch complex signals to obtain the signal energy statistics corresponding to the N sampling points. Considering the resource consumption, the normalization buffer module will also be called;

[0148] The timing result calculation module uses the bit timing estimation algorithm of trigonometric interpolation to estimate that the deviation distribution approximately follows the sine distribution, expands it into a Fourier series containing only sine terms and ignores high-order harmonics to correct the estimation result; the angle calculation in the project is implemented by the CORDIC module, and the calculation formula is implemented by storing the result in the ROM and looking up the table;

[0149] The timing error correction module selects the signal sampling moment corresponding to the bit timing result through the FARROW interpolation filter, and finally obtains the signal corresponding to the best sampling moment.

[0150] According to one aspect of the present application, a bit timing estimation device suitable for satellite communication includes:

[0151] at least one processor; and,

[0152] a memory communicatively connected to at least one of the processors; wherein,

[0153] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the bit timing estimation method applicable to satellite communication described in any one of the above embodiments.

[0154] The present invention achieves high accuracy and high reliability of timing estimation in satellite communications by constructing a multi-level and multi-dimensional signal processing and feature extraction system. From feature extraction, fine estimation, performance evaluation to parameter optimization, a complete closed-loop optimization system is formed. Feature space mapping and nonlinear transformation provide a comprehensive characterization of signal features, adaptive filtering and state tracking ensure the accuracy of the estimation results, multi-dimensional performance evaluation and real-time monitoring ensure the stable operation of the system, and feature fusion and decision optimization achieve continuous optimization of parameters. The present invention improves the synchronization performance of satellite communication systems.

[0155] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A bit timing estimation method suitable for satellite communications, characterized in that: The steps include: S1. Obtain an original complex baseband signal, map the original complex baseband signal to a four-dimensional feature space, and obtain a feature matrix; Perform piecewise nonlinear transformation and dimensionality reduction processing on the feature matrix to obtain the transformed feature matrix; use a recursive sliding window to process the transformed feature matrix and extract the time synchronization feature vector; based on the time synchronization feature vector, perform multi-layer adaptive fusion operation to obtain the initial timing point estimation value; S2. Based on the initial timing point estimation value, extract the received signal of the corresponding time window, and input it into the adaptive filter group for processing to obtain the time domain feature vector, the frequency domain feature vector and the phase feature vector; process the time domain feature vector, the frequency domain feature vector and the phase feature vector through a nonlinear mapping network to obtain an adaptive weight vector; based on the adaptive weight vector, perform fusion calculation and interpolation optimization on the time domain feature vector, the frequency domain feature vector and the phase feature vector to obtain a fine timing estimation value; Based on the precise timing estimate, a state space model is constructed to perform dynamic tracking and obtain the final timing correction value. S3, performing multi-dimensional performance evaluation on the final timing correction value and the pre-stored historical timing sequence to obtain a performance index vector; based on the performance index vector, performing anti-interference processing on the received signal to obtain an anti-interference processed signal; Obtain current system parameters, perform optimization calculation based on the current system parameters and performance indicator vector, and obtain an optimized parameter set; S4, based on the anti-interference processing signal and the optimized parameter set, perform time-frequency joint analysis and phase space reconstruction to obtain a fusion feature vector; based on Fuse feature vectors, perform multi-level decision operations, and obtain decision results; Based on the decision results and pre-stored historical performance data, the system is self-calibrated to obtain a calibration parameter set; the calibration parameter set and pre-stored historical data are input into the prediction optimization module to obtain the final optimization parameters and performance prediction report; Step S1 is further as follows: S11, obtaining an original complex baseband signal, processing the original complex baseband signal through an amplitude feature extraction function to obtain a signal amplitude feature vector; processing the original complex baseband signal through a phase feature extraction function to obtain a phase jump feature vector; processing the original complex baseband signal through a frequency feature extraction function to obtain a frequency offset feature vector; processing the original complex baseband signal through a time feature extraction function to obtain a time-related feature vector; combining the signal amplitude feature vector, the phase jump feature vector, the frequency offset feature vector and the time-related feature vector to obtain a feature matrix; S12, calculating the norm value of the characteristic matrix to obtain a norm determination result; based on the norm determination result, selecting a corresponding nonlinear transformation function; Based on the nonlinear transformation function, the characteristic matrix is ​​processed to obtain the transformed characteristic matrix; S13, based on the pre-stored signal-to-noise ratio parameter and time correlation parameter, calculate the window size value; based on the window size value, recursively slide the window processing on the transformed feature matrix to obtain a window feature sequence; calculate the weighted feature combination value, the autocorrelation function envelope peak position and the feature difference of each window in the window feature sequence; combine the weighted feature combination value, the autocorrelation function envelope peak position and the feature difference to obtain a time synchronization feature vector; S14, performing exponential weighted moving average filtering on the time synchronization feature vector to obtain a first-level filtering result; multiplying the first-level filtering result by a preset adaptive weight matrix to obtain a feature weight; The feature weights are weighted with the first-level filtering results, and the time point corresponding to the maximum value is taken to obtain the initial timing point estimate.

2. The bit timing estimation method suitable for satellite communication according to claim 1, characterized in that: Step S2 is further as follows: S21, based on the initial timing point estimation value, extract the received signal of the corresponding time window; input the received signal into the adaptive bandpass filter for processing to obtain the time domain feature vector; process the received signal through the improved Hilbert transform to obtain the frequency domain feature vector; calculate the conjugate product of the frequency domain feature vector and the time domain feature vector, and obtain the phase feature vector after phase unwrapping processing; S22, performing a first-level nonlinear mapping based on the time domain feature vector, the frequency domain feature vector and the phase feature vector to obtain a first-level mapping result; performing a second-level nonlinear mapping based on the first-level mapping result to obtain a second-level mapping result; Based on the secondary mapping results, the third-level nonlinear mapping and softmax operation are performed to obtain an adaptive weight vector; S23, based on the adaptive weight vector, weighted fusion is performed on the time domain feature vector, the frequency domain feature vector and the phase feature vector to obtain a fused feature vector; The fused feature vectors are processed by cubic spline interpolation to obtain an interpolation optimization result; an error compensation value is calculated based on the interpolation optimization result and pre-stored channel state parameters; the interpolation optimization result and the error compensation value are added to obtain a fine timing estimation value; S24, constructing a state vector based on the fine timing estimate, including timing position, velocity and acceleration components; predicting the state vector using a state transfer matrix to obtain a state prediction value; calculating a Kalman gain based on the state prediction value and the fine timing estimate to obtain a gain matrix; Based on the gain matrix, the state vector is corrected to obtain the final timing correction value.

3. The bit timing estimation method suitable for satellite communication according to claim 2, characterized in that: Step S3 is further as follows: S31, calculating the weighted difference between the final timing correction value and the pre-stored historical correction value to obtain a steady-state performance index; performing exponential weighted calculation on the variance of the final timing correction value based on the pre-stored signal-to-noise ratio parameter to obtain a dynamic performance index; using the Monte Carlo method to evaluate the reliability of the final timing correction value to obtain a reliability index; combining the steady-state performance index, the dynamic performance index and the reliability index to obtain a performance index vector; S32, based on the performance indicator vector, performing time-frequency transformation processing on the received signal to obtain a time-frequency distribution diagram; calculating the peak-to-average ratio of the time-frequency distribution diagram to obtain an interference indicator value; based on the interference indicator value, constructing adaptive filter parameters, filtering the received signal, and obtaining a filtered signal; Processing the filtered signal through a nonlinear suppression function to obtain an anti-interference processed signal; S33, obtaining current system parameters, and calculating the objective function value based on the current system parameters and the performance indicator vector; Based on the gradient information of the objective function value, the system parameters are updated to obtain an updated parameter set; Project the updated parameter set to obtain the optimized parameter set; S34, constructing a performance prediction model based on the optimization parameter set and the pre-stored environmental parameters to obtain a performance prediction value; Calculate the deviation between the performance prediction and the actual performance to obtain the anomaly detection statistic; Based on anomaly detection statistics, determine whether parameter re-optimization needs to be triggered and generate parameter adjustment strategies; Based on the parameter adjustment strategy, the system parameters are adjusted to obtain monitoring adjustment parameters; Based on the monitoring, the parameters are adjusted and the optimization parameter set is updated to obtain an updated optimization parameter set.

4. The bit timing estimation method suitable for satellite communication according to claim 3, characterized in that: Step S4 is further as follows: S41. Based on the anti-interference processing signal and the optimized parameter set, an improved Wigner-Ville distribution calculation is performed to obtain a time-frequency distribution result; based on the time-frequency distribution result, a delay matrix of the anti-interference processing signal is constructed, and phase space reconstruction is performed to obtain a reconstructed feature matrix; the time-frequency distribution result and the reconstructed feature matrix are processed by tensor fusion operation to obtain a fused feature vector; S42, based on the fused feature vector, performing a first-level decision operation to obtain a pattern recognition result; based on the pattern recognition result, selecting parameters from the optimization parameter set and further optimizing them to obtain an optimization parameter group; Calculate the reliability evaluation value of the pattern recognition result and the optimized parameter group to obtain the decision result; S43, based on the decision result and the pre-stored historical performance data, performing error analysis to obtain the systematic error and random error; Perform exponential decay compensation on the system error to obtain the system compensation value; Perform median filtering on the random error to obtain a random compensation value; based on the system compensation value and the random compensation value, update the system parameters to obtain a calibration parameter set; S44, constructing a time series model based on pre-stored historical performance data to obtain a performance trend value; Based on the performance trend value and the calibration parameter set, the risk assessment function value is calculated to obtain the risk assessment result; The risk assessment result is input into the feedback controller to adjust the parameters and obtain the control adjustment amount; Add the calibration parameter set and control adjustments to obtain the final optimized parameters and performance prediction report.

5. The bit timing estimation method suitable for satellite communication according to claim 4, characterized in that: Step S11 is further as follows: S111, obtaining an original complex baseband signal, performing adaptive piecewise linear quantization processing to obtain a normalized signal; performing a multi-scale wavelet transform on the normalized signal to obtain a multi-scale coefficient matrix; Based on the multi-scale coefficient matrix, the energy distribution of each scale is calculated to obtain the energy distribution vector; the energy distribution vector is input into the chaotic mapping unit for feature enhancement to obtain the signal amplitude feature vector; S112, the original complex baseband signal is subjected to Hilbert transformation to obtain an analytical signal; the instantaneous phase of the analytical signal is calculated to obtain a phase sequence; the phase sequence is subjected to adaptive threshold detection to obtain a phase jump point; based on the phase jump point, a local phase change feature is extracted to obtain a local phase feature; the local phase feature is compressed through a neural network encoder to obtain a phase jump feature vector; S113, performing segmented windowing processing on the original complex baseband signal to obtain a first segmented signal sequence; performing adaptive fractional Fourier transform on the first segmented signal sequence to obtain a spectrum sequence; calculating the center of gravity drift of the spectrum sequence to obtain a frequency drift sequence; performing fuzzy clustering processing on the frequency drift sequence to obtain a frequency offset feature vector; S114, performing autocorrelation calculation based on the original complex baseband signal to obtain an autocorrelation sequence; performing nonlinear feature extraction on the autocorrelation sequence to obtain a nonlinear feature sequence; Calculate the recursive quantization entropy of the nonlinear feature sequence to obtain an entropy value sequence; process the entropy value sequence through a deep residual network to obtain a time-dependent feature vector; S115, performing feature fusion based on the signal amplitude feature vector, the phase jump feature vector, the frequency offset feature vector and the time correlation feature vector to obtain an initial feature combination; performing adaptive weight calculation on the initial feature combination to obtain a feature weight vector; Based on the feature weight vector, the initial feature combination is weighted and combined to obtain the feature matrix.

6. The bit timing estimation method suitable for satellite communication according to claim 4, characterized in that: Step S12 is further as follows: S121, calculating the first-class norm value and the second-class norm value of the feature matrix, dividing the difference between the first-class norm value and the second-class norm value by the second-class norm value to obtain a norm ratio; filtering the norm ratio using an exponential moving average method to obtain a smoothed norm ratio sequence; calculating the statistical moment of the norm ratio sequence using a dynamic time window; determining a norm determination threshold based on the distribution characteristics of the statistical moment; S122, comparing the current norm ratio with the norm determination threshold to obtain a norm interval identifier; based on the norm interval identifier, selecting a corresponding transformation function from a preset nonlinear function set; using an adaptive parameter estimation method to calculate a shape parameter and a scaling parameter of the transformation function; combining the transformation function and its shape parameter and scaling parameter to form a complete nonlinear transformation scheme; S123, based on the nonlinear transformation scheme, performing local response normalization preprocessing on the feature matrix, calculating the local root mean square value of the feature matrix; dividing each element of the feature matrix by the corresponding local root mean square value to obtain a normalized feature matrix; A variational autoencoder is used to extract the latent variable representation of the normalized feature matrix; the latent variable representation is sparsely regularized to obtain a compressed feature representation; S124, construct a multi-head attention processing unit, and input the compressed feature representation into a predetermined number of parallel attention subspaces respectively; Calculate the query matrix, key matrix and value matrix in each attention subspace; multiply the query matrix by the transpose of the key matrix and scale normalize to get the attention score; Multiply the attention score with the value matrix to get the attention output of each subspace; concatenate the attention outputs of all subspaces and perform linear transformation to get the fused feature matrix; S125, applying the transformation function to the fusion feature matrix to obtain a preliminary transformation result; The residual connection method is used to fuse the initial transformation results with the feature matrix to obtain the fused features; Perform layer normalization on the fused features to obtain normalized features; The position coding modulation normalization features of adaptive learning are used to obtain the final transformed feature matrix, that is, the transformed feature matrix.

7. The bit timing estimation method suitable for satellite communication according to claim 4, characterized in that: Step S21 is further as follows: S211, constructing an adaptive piecewise function based on the initial timing point estimate to obtain a window boundary sequence; extracting a received signal corresponding to the time window based on the initial timing point estimate; segmenting the received signal according to the window boundary sequence to obtain a second segmented signal sequence; performing dynamic overlapping processing on the second segmented signal sequence to obtain an overlapping signal sequence; performing spline interpolation on the overlapping signal sequence to obtain an interpolated signal sequence; S212, input the interpolation signal sequence into the adaptive filter group to obtain a multi-channel filtering result; perform envelope extraction on the multi-channel filtering result to obtain an envelope sequence group; calculate the cross-correlation function of the envelope sequence group to obtain a related feature sequence; The relevant feature sequence is processed by a deep feature extraction network to obtain a time domain feature vector; S213, performing wavelet packet decomposition on the interpolation signal sequence to obtain a time-frequency coefficient matrix; extracting energy distribution characteristics of the time-frequency coefficient matrix to obtain a first energy distribution diagram; inputting the first energy distribution diagram into an improved Hilbert transformer to obtain a transformation coefficient sequence; performing principal component analysis on the transformation coefficient sequence to obtain a frequency domain feature vector; S214, processing the time domain feature vector and the frequency domain feature vector through a tensor network to obtain a fusion feature; The local phase information of the fused features is calculated to obtain a phase sequence; the phase sequence is adaptively deconvolved to obtain a deconvolution result; the deconvolution result is processed through a nonlinear mapping network to obtain a phase feature vector.

8. The bit timing estimation method suitable for satellite communication according to claim 4, characterized in that: Step S32 is further as follows: S321, based on the performance indicator vector, adaptively segment the received signal to obtain a third segment signal sequence; apply an improved Cohen-like distribution transform to the third segment signal sequence to obtain a time-frequency distribution matrix; Calculating the local energy concentration of the time-frequency distribution matrix to obtain a second energy distribution map; The second energy distribution graph is processed by hyperbolic shear transformation to obtain a time-frequency feature matrix; S322, perform singular value decomposition on the time-frequency feature matrix to obtain a singular value sequence; calculate the statistical moment of the singular value sequence to obtain a statistical feature vector; The statistical feature vector is input into the fuzzy inference system to obtain the interference feature vector; based on the interference feature vector, the interference measurement value is calculated to obtain the interference index value; S323, constructing an adaptive basis function based on the interference index value to obtain a filter basis group; performing orthogonalization processing on the filter basis group to obtain an orthogonal basis function group; calculating the optimal combination coefficient of the orthogonal basis function group to obtain filter parameters; Based on the filter parameters, an adaptive filter is constructed to filter the received signal to obtain a filtered signal; S324, performing wavelet threshold processing based on the filtered signal to obtain a threshold processing result; Morphological filtering is performed on the threshold processing result to obtain morphological features; the morphological features are processed by a nonlinear suppression function to obtain a suppression result; and adaptive gain control is performed on the suppression result to obtain an anti-interference processing signal.

9. A bit timing estimation device suitable for satellite communication, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the bit timing estimation method applicable to satellite communication as described in any one of claims 1 to 8.

Images