A range search type Doppler frequency offset estimation system for a high-speed moving scene

Abstract

The application discloses a kind of interval search formula Doppler frequency deviation estimation systems for high-speed moving scene, it is related to wireless communication technical field, including: channel perception module, the channel perception module is used to receive fast time-varying channel parameter and terminal motion parameter, adaptive multistage interval contraction search module, frequency deviation ambiguity decoupling module, frequency deviation fine calibration module, closed loop parameter control module, the closed loop parameter control module is used to receive global frequency deviation estimation value, difference calculation is carried out to global frequency deviation estimation value, and frequency deviation jump rate is obtained, according to jump rate adaptive adjustment search interval width and search layer number;By adaptive multistage interval contraction search module, subinterval division granularity and contraction algorithm are adaptively adjusted in combination with channel signal-to-noise ratio;While through closed loop parameter control module, search parameter is adaptively adjusted based on the jump rate of global frequency deviation estimation value, forms complete parameter closed loop control mechanism.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically an interval search-type Doppler frequency offset estimation system for high-speed mobile scenarios. Background Technology

[0002] With the evolution of mobile communication technology, the application demand for high-speed mobile edge communication scenarios such as high-speed rail, low-altitude drone communication, and high-dynamic vehicle-mounted networks is growing rapidly. In these scenarios, the high-speed movement of terminals generates large dynamic and rapidly changing Doppler frequency offsets, and under extreme conditions, spectrum folding and blurring may occur. At the same time, the wireless channel exhibits fast time-varying and deep fading characteristics, which directly leads to a sharp drop in the signal demodulation performance of the receiving end. This is the core bottleneck for ensuring high-reliability and low-latency transmission of high-speed edge communication links.

[0003] Existing Doppler frequency offset estimation techniques suffer from the following core technical defects directly related to this solution in practical applications of high-speed mobile edge communication scenarios:

[0004] The frequency offset search range and search strategy are fixed and cannot be dynamically adapted by combining the time-varying characteristics of the channel, the real-time motion state of the terminal and the channel signal-to-noise ratio. It cannot simultaneously take into account the convergence speed of frequency offset search and the coverage capability of large dynamic frequency offset. Furthermore, it is prone to estimation loss in the scenario of frequency offset change caused by high-speed movement, and cannot take into account the real-time performance and robustness of the solution.

[0005] To address the issue of folding and blurring of integer multiple sampling frequency offsets in high-speed, large-frequency-off scenarios, there is a lack of effective identification and decoupling correction mechanisms, which can easily introduce integer frequency offset estimation errors and fail to guarantee estimation accuracy under large dynamic frequency offset conditions.

[0006] Existing solutions are mostly designed for high-computing-power equipment at the base station end, without taking into account the core needs of edge terminals with limited computing power and local data privacy protection. At the same time, they lack a long-term parameter drift calibration mechanism, making it impossible to achieve high-precision, high-stability, and high-security large-scale deployment on low-computing-power edge terminals. Summary of the Invention

[0007] The purpose of this invention is to propose an interval-search Doppler frequency offset estimation system for high-speed moving scenarios, comprising:

[0008] The channel sensing module is used to receive fast time-varying channel parameters and terminal motion parameters, perform autocorrelation calculation on the received channel impulse response to obtain the channel coherence time threshold, calculate the maximum theoretical frequency offset at the current motion speed according to the Doppler formula, eliminate invalid frequency offset intervals that exceed the range, and output the initial frequency offset search interval.

[0009] An adaptive multi-level interval shrinkage search module is used to receive the initial search interval and closed-loop control parameters. First, it detects the signal-to-noise ratio of the current channel, and then adaptively adjusts the division granularity and shrinkage algorithm of the sub-interval based on the signal-to-noise ratio. Finally, it uses the maximum allowable frequency offset change within the channel coherence window as the convergence threshold to determine search convergence and outputs a coarse frequency offset estimation result.

[0010] The frequency offset fuzzy decoupling module is used to receive the frequency offset coarse estimation result, perform differential detection on the frequency offset coarse estimation result for a consecutive preset number of frames, identify the folding factor of the sampling frequency offset as an integer multiple, correct the coarse estimation result according to the detected folding factor, and output the decoupled frequency offset result.

[0011] The frequency offset fine-tuning module is used to receive the decoupled frequency offset result, perform singular value decomposition on the decoupled frequency offset result, project the result onto the signal subspace to remove noise components, perform differential detection to mark the jump frame on the decoupled frequency offset result for a consecutive preset number of frames, generate frequency offset confidence weight based on the channel correlation coefficient and the jump mark, and output the frequency offset confidence weight.

[0012] The closed-loop parameter control module receives the global frequency offset estimate, performs differential calculation on the global frequency offset estimate to obtain the frequency offset jump rate, adaptively adjusts the search interval width and search layer number according to the jump rate, performs linear interpolation adjustment on the parameters corresponding to the intermediate rate, and outputs the control parameters for the interval search, forming a closed loop with the adaptive multi-level interval shrinkage search module.

[0013] Furthermore, it also includes a space-time-frequency preprocessing module, which is used to perform space-time joint filtering on the received baseband signal. It adopts an adaptive filtering window of a preset size, and the weight of the filter is adaptively adjusted according to the delay spread of the current channel to separate short-delay multipath coupling components. Then, the filtered signal is cross-correlated with the local synchronization preamble symbol to complete symbol-level alignment, and the alignment error is controlled within a set range. Finally, frequency domain noise whitening processing is performed.

[0014] Furthermore, the channel sensing module is used to first perform autocorrelation calculation on the received channel impulse response to obtain the channel coherence time threshold, and then collect the terminal's moving speed, three-axis acceleration and heading angle parameters at a preset sampling rate through the terminal's built-in IMU inertial measurement unit. Finally, it calculates the maximum theoretical frequency offset at the current moving speed according to the Doppler formula, eliminates invalid frequency offset intervals that exceed the range, and determines the initial search interval.

[0015] Furthermore, it also includes an interval search module. The interval search module is used to first detect the signal-to-noise ratio (SNR) of the current channel and set a preset threshold. When the channel SNR is higher than the preset threshold, the initial search interval is divided into sub-intervals with a preset granularity. The golden section method is used to perform sub-interval shrinkage, and the iteration step size is set to a preset value. When the channel SNR is not higher than the preset threshold, the initial search interval is divided into sub-intervals with a preset granularity and the bisection method is used to perform sub-interval shrinkage, with the iteration step size set to a preset value. Finally, the maximum allowable frequency offset change within the channel coherence window is used as the convergence threshold to determine search convergence and output the coarse frequency offset estimation result.

[0016] Furthermore, the frequency offset fuzzy decoupling module is used to first perform differential detection on the frequency offset coarse estimation result of a consecutive preset number of frames. When the magnitude of the differential result is greater than half of the sampling frequency, it is determined that there is frequency offset folding. The corresponding folding factor is the rounded result of the differential result divided by the sampling frequency. Then, the coarse estimation result is corrected according to the detected folding factor.

[0017] Furthermore, the frequency offset fine-tuning module is used to first perform singular value decomposition on the covariance matrix composed of the decoupled frequency offset results, project the decomposed results onto the signal subspace corresponding to the first K largest singular values, remove the interference components of the noise subspace, then perform differential detection on the decoupled frequency offset results for a consecutive preset number of frames, use a preset value as a preset jump threshold, when the differential result is greater than the threshold, mark the frame as a jump frame, and finally generate a frequency offset confidence weight ranging from 0 to 1 based on the channel correlation coefficient and the jump mark.

[0018] Furthermore, it also includes a multi-antenna federated fusion and privacy protection module, which is used to first perform scale normalization on the frequency offset characteristics of each antenna to eliminate the feature deviation of heterogeneous channels.

[0019] Then, the fusion weights are allocated according to the frequency offset confidence weights. The antenna weights with confidence values ​​lower than the preset value are reset to 0. Finally, the local features are distilled into global features of preset dimensions through a two-layer fully connected network. At the same time, Laplacian noise is added to the local features. The noise scale is the feature sensitivity divided by the privacy budget to protect the local privacy of multiple antennas.

[0020] Furthermore, the closed-loop parameter control module is used to first perform differential calculation on the global frequency offset estimate to obtain the frequency offset jump rate, and use a preset value and a preset value as the jump rate threshold. When the jump rate is greater than the preset value, the search interval width is adjusted to the preset value and the search layer is increased to the preset layer. When the jump rate is less than or equal to the preset value, the search interval width is adjusted to the preset value and the search layer is reduced to the preset layer. The parameters corresponding to the intermediate rate are linearly interpolated and adjusted to form a closed loop with the interval search module.

[0021] Furthermore, it also includes a lightweight edge inference module. The lightweight edge inference module is used to first perform adaptive low-bit quantization on the interval search model, quantize the parameters of FP32 to the precision of INT8, and control the quantization error within a reasonable range obtained based on calibration data. Then, it eliminates low-priority search paths with confidence values ​​lower than preset values ​​based on frequency offset confidence weights. Finally, it adaptively adjusts the inference precision based on the computing power level of the edge terminal to adapt to terminals with different computing power.

[0022] Furthermore, it also includes an anomaly calibration module, which first uses a sliding window of a preset length to perform a Grubbs statistical test on the global frequency offset estimate within the window to remove abnormal data that deviate from the confidence interval of the mean plus or minus a preset multiple of the standard deviation. Then, based on the effective frequency offset data after removing anomalies, it uses a mini-batch stochastic gradient descent method with a preset learning rate to incrementally update the calibration parameters of the initial search interval.

[0023] The technical solution of the present invention brings at least the following beneficial effects:

[0024] By employing an adaptive multi-level interval shrinkage search module, combined with an adaptive adjustment of the sub-interval division granularity and shrinkage algorithm based on the channel signal-to-noise ratio, and a closed-loop parameter control module that adaptively adjusts the search parameters based on the jump rate of the global frequency offset estimate, a complete parameter closed-loop control mechanism is formed. This significantly reduces the frequency offset search range and improves the convergence speed, while avoiding the problem of missing detection of large frequency offsets in fixed intervals. It can still ensure the stability of the estimation in frequency offset change scenarios, perfectly adapting to the fast time-varying channel characteristics in high-speed mobile scenarios.

[0025] By performing differential detection on the coarse frequency offset estimation results of multiple consecutive frames through the frequency offset fuzzy decoupling module, the folding factor of the integer multiple sampling frequency offset is accurately identified, and the coarse estimation results are corrected and decoupled. This effectively solves the spectral folding fuzzy problem in high-speed and large frequency offset scenarios, eliminates integer frequency offset estimation error, and improves the estimation accuracy under large dynamic frequency offset conditions.

[0026] By employing federated knowledge distillation and differential privacy mechanisms, the system balances the accuracy of multi-antenna fusion with the local privacy and security of the terminal. Through a lightweight inference module at the edge, the system achieves lightweight deployment of a high-precision frequency offset estimation scheme on edge terminals with low computing power. Simultaneously, through an anomaly calibration module, the system accurately identifies and removes abnormal data, incrementally updates system calibration parameters, corrects the parameter drift problem that has been running for a long time, and improves the stability and adaptability of the system for continuous operation. Attached Figure Description

[0027] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0029] Please see Figure 1 This application provides an interval-search Doppler frequency offset estimation system for high-speed moving scenarios, comprising:

[0030] The channel sensing module is used to receive fast time-varying channel parameters and terminal motion parameters. The fast time-varying channel parameters are obtained through least squares channel measurement, and the terminal motion parameters are obtained through an IMU inertial measurement unit. The module performs autocorrelation calculation on the received channel impulse response to obtain the channel coherence time threshold, calculates the maximum theoretical frequency offset at the current motion speed according to the Doppler formula, eliminates invalid frequency offset intervals that exceed the range, and outputs the initial frequency offset search interval.

[0031] An adaptive multi-level interval shrinkage search module is used to receive the initial search interval and closed-loop control parameters. First, it detects the signal-to-noise ratio of the current channel, and then adaptively adjusts the division granularity and shrinkage algorithm of the sub-interval based on the signal-to-noise ratio. Finally, it uses the maximum allowable frequency offset change within the channel coherence window as the convergence threshold to determine search convergence and outputs a coarse frequency offset estimation result.

[0032] The frequency offset fuzzy decoupling module is used to receive the frequency offset coarse estimation result, perform differential detection on the frequency offset coarse estimation result for a consecutive preset number of frames, identify the folding factor of the sampling frequency offset as an integer multiple, correct the coarse estimation result according to the detected folding factor, and output the decoupled frequency offset result.

[0033] The frequency offset fine-tuning module is used to receive the decoupled frequency offset result, perform singular value decomposition on the decoupled frequency offset result, project the result onto the signal subspace to remove noise components, perform differential detection to mark the jump frame on the decoupled frequency offset result for a consecutive preset number of frames, generate frequency offset confidence weight based on the channel correlation coefficient and the jump mark, and output the frequency offset confidence weight.

[0034] The closed-loop parameter control module receives the global frequency offset estimate, performs differential calculation on the global frequency offset estimate to obtain the frequency offset jump rate, adaptively adjusts the search interval width and search layer number according to the jump rate, performs linear interpolation adjustment on the parameters corresponding to the intermediate rate, and outputs the control parameters for the interval search, forming a closed loop with the adaptive multi-level interval shrinkage search module.

[0035] It should be noted that the fast time-varying channel parameters are obtained by performing least-squares channel measurements on the received pilot symbols. Least-squares channel measurement refers to the measurement method that estimates the impulse response of the channel by minimizing the reception error of the pilot symbols.

[0036] Terminal motion parameters are calculated through acceleration integration and angular velocity integration of the IMU, where the IMU refers to the inertial sensor built into the terminal for measuring motion parameters; fast time-varying channel parameters refer to channel coherence time, channel correlation coefficient, and delay spread parameters, where channel coherence time refers to the maximum length of time for the channel impulse response to remain stable, channel correlation coefficient refers to the correlation between the channel impulse responses of adjacent frames, and delay spread refers to the standard deviation of the delay of multipath signals;

[0037] Terminal motion parameters refer to the terminal's moving speed, three-axis acceleration, and heading angle. The moving speed refers to the terminal's moving rate relative to the base station, the three-axis acceleration refers to the terminal's acceleration in three spatial dimensions, and the heading angle refers to the angle between the terminal's direction of motion and the base station's direction. Autocorrelation calculation refers to the method for calculating the correlation between the channel impulse response and its own delayed version. The channel coherence time threshold refers to the time length corresponding to when the autocorrelation coefficient drops to a set value, which is the definition of coherence time for Rayleigh fast fading channels.

[0038] The Doppler formula is a physical formula describing the relationship between the Doppler frequency offset and the motion speed of a mobile terminal; the initial search interval refers to the initial range of subsequent interval searches, used to narrow the search range; the closed-loop control parameters refer to the interval width and search layer number parameters generated based on the global frequency offset jump rate, where the interval width refers to the frequency range size of the search interval, and the search layer number refers to the number of iterations for interval shrinkage; the signal-to-noise ratio is the ratio of signal power to noise power; the granularity of sub-interval division refers to the frequency range size of each sub-interval, and the shrinkage algorithm refers to the iterative algorithm used to narrow the search interval;

[0039] The channel coherence window refers to a time window with the channel coherence time as its length. The maximum allowable frequency offset change refers to the maximum possible change in frequency offset within this window. The convergence threshold is the threshold used to determine whether the search has converged. The coarse frequency offset estimate refers to the preliminary frequency offset estimate obtained from the interval search. The number of consecutive preset frames refers to a preset number of data frames that are consecutive in time. This number of frames is determined by both the channel coherence time and the frame interval to ensure the validity of the differential results (in this scheme, the preset number of frames is 5, which comes from the adaptation calculation of the channel coherence time and the frame interval to ensure the accuracy of differential detection).

[0040] Differential detection refers to a detection method that calculates the difference between the frequency offset results of adjacent frames; folding of integer multiple sampling frequency offset refers to the phenomenon that when the Doppler frequency offset exceeds half of the baseband sampling frequency, the frequency offset will be folded into the sampling frequency range; the folding factor refers to the integer coefficient used to correct the folded frequency offset; the decoupled frequency offset result refers to the unambiguous frequency offset result after folding correction; the covariance matrix refers to the matrix used to describe the statistical correlation of the frequency offset result; singular value decomposition refers to the matrix decomposition method that decomposes a matrix into a left singular matrix, a singular value matrix, and a right singular matrix, used to separate signal and noise components;

[0041] The signal subspace refers to the subspace composed of eigenvectors corresponding to large singular values, used to represent signal components; the noise subspace refers to the subspace composed of eigenvectors corresponding to small singular values, used to represent noise components; a jump frame refers to a frame where the frequency offset changes abruptly; confidence weight refers to the weight used to represent the reliability of the frequency offset result, ranging from 0 to 1; the global frequency offset estimate refers to the final frequency offset estimate result after multi-antenna fusion; differential calculation refers to the method of calculating the derivative of the frequency offset with respect to time; the frequency offset jump rate refers to the change in frequency offset per unit time; linear interpolation refers to the interpolation method of calculating the value of the intermediate point using the values ​​of two known points; closed loop refers to using the output of the control module as the input of the search module to achieve adaptive adjustment of parameters.

[0042] As an optional embodiment, it also includes a space-time-frequency preprocessing module, which is used to perform space-time joint filtering on the received baseband signal, adopts an adaptive filtering window of preset size, and the weight of the filter is adaptively adjusted according to the delay spread of the current channel to separate short-delay multipath coupling components. Then, the filtered signal is cross-correlated with the local synchronization preamble symbol to complete symbol-level alignment, and the alignment error is controlled within a set range. Finally, frequency domain noise whitening processing is performed.

[0043] Space-time joint filtering refers to a joint filtering method that is executed simultaneously in the spatial and temporal dimensions to suppress multipath interference. An adaptive filtering window of preset size (a 3×3 window is used in this scheme, derived from the adaptation calculation of the system sampling rate and multipath delay characteristics; the 3x3 window, with rows corresponding to the 3 sampling points in the time dimension and columns corresponding to the 3 antenna receiving channels in the spatial dimension, can precisely cover the delay range of short-delay multipath in high-speed scenarios); filter weights refer to the filtering coefficients used to perform weighted summation of the input signals.

[0044] Short-delay multipath coupling components refer to multipath signal components with delays less than a set value. These components can cause interference to the signal in high-speed scenarios. Cross-correlation refers to the calculation method for the correlation between two signals. Synchronization preamble refers to the preamble symbol stored locally for synchronization. Symbol-level alignment refers to the process of aligning the received signal with the local symbol.

[0045] Alignment error refers to the time error after alignment; symbol period refers to the duration of a baseband symbol; frequency domain noise whitening is a processing method that converts colored noise into white noise, used to suppress colored noise caused by fast fading channels.

[0046] The maximum latency of short-latency multipath in high-speed mobile scenarios is a set value, and the baseband sampling rate of this system is a set value, and the corresponding sampling interval is a set value. Therefore, the window of the preset size can just cover this latency range. The size of the window is determined by the system's sampling rate and multipath latency characteristics.

[0047] The calculation steps for frequency domain noise whitening are as follows: the pre-whitening matrix W is equal to the negative 1 / 2 power of R, where R is the noise covariance matrix of the channel, which is obtained through the statistics of the preamble symbol.

[0048] As an optional embodiment, the channel sensing module is used to first perform autocorrelation calculation on the received channel impulse response to obtain the channel coherence time threshold, and then collect the terminal's moving speed, three-axis acceleration and heading angle parameters at a preset sampling rate through the terminal's built-in IMU inertial measurement unit. Finally, it calculates the maximum theoretical frequency offset at the current moving speed according to the Doppler formula, eliminates invalid frequency offset intervals that exceed the range, and determines the initial search interval.

[0049] Channel impulse response refers to the impulse response of a channel, describing the channel's response to a pulse signal. Autocorrelation calculation is a method for calculating the correlation between the channel impulse response and its own delayed version. Channel coherence time threshold refers to the time length corresponding to when the autocorrelation coefficient drops to a set value, which is the definition of coherence time for Rayleigh fast fading channels.

[0050] An IMU (Inertial Measurement Unit) refers to an inertial sensor built into the terminal for measuring motion parameters. The sampling rate refers to the number of times data is collected per second. The preset sampling rate (100Hz in this solution, derived from the calculation of parameter update requirements in high-speed mobile scenarios, which can ensure that the update speed of motion parameters meets the parameter update requirements in high-speed mobile scenarios)

[0051] The Doppler formula is a physical formula describing the relationship between the Doppler frequency offset and the motion speed of a mobile terminal. Invalid frequency offset intervals outside the range are eliminated to determine the initial search interval. The initial search interval refers to the initial range for subsequent interval searches, used to narrow the search range. The calculation steps for channel coherence time are as follows: Autocorrelation is performed on the channel impulse response h(τ). The autocorrelation Rh(Δt) of the channel impulse response is equal to the expectation of h(t) multiplied by the conjugate of h(t minus Δt). When Rh(Δt) drops to a set value, the corresponding Δt is the coherence time threshold. This threshold is the definition of coherence time for Rayleigh fast fading channels, where h(t) is the channel impulse response at time t, and Δt is the time difference.

[0052] The calculation steps of the Doppler formula are as follows: the Doppler frequency offset fd is equal to the terminal moving speed v multiplied by the carrier frequency fc multiplied by the cosine value of the angle between the moving directions cosθ, and then divided by the speed of light c, where v is the terminal moving speed, fc is the carrier frequency, θ is the angle between the terminal moving direction and the base station direction, and c is the speed of light. All parameters of this formula come from the motion parameters collected in real time by the IMU. The maximum theoretical frequency offset fmax is calculated by this formula, and the initial search interval is set to negative fmax to positive fmax.

[0053] As an optional embodiment, an interval search module is also included. The interval search module is used to first detect the signal-to-noise ratio (SNR) of the current channel, and set a preset value as a preset threshold. When the channel SNR is higher than the preset threshold, the initial search interval is divided into sub-intervals with a preset granularity. The golden section method is used to perform sub-interval shrinkage, and the iteration step size is set to a preset value. When the channel SNR is not higher than the preset threshold, the initial search interval is divided into sub-intervals with a preset granularity, and the bisection method is used to perform sub-interval shrinkage, with the iteration step size set to a preset value. Finally, the maximum allowable frequency offset change within the channel coherence window is used as the convergence threshold to determine search convergence, and the frequency offset coarse estimation result is output.

[0054] Signal-to-noise ratio (SNR) is the ratio of signal power to noise power. The preset threshold (in this scheme, it is set to 15dB, which is derived from the performance crossover point measurement of the golden section method and the bisection method in high-speed scenarios. When the SNR is higher than this threshold, the convergence speed of the golden section method is better than that of the bisection method. When the SNR is lower than this threshold, the noise resistance of the bisection method is better than that of the golden section method).

[0055] The granularity of sub-interval division refers to the frequency range of each sub-interval. The preset granularity is 50Hz under high signal-to-noise ratio (SNR), which is derived from the noise level measurement under high SNR. Under high SNR, the noise is small, so the sub-interval can be smaller.

[0056] The golden section method is an iterative search algorithm that repeatedly divides the search interval according to the golden ratio and eliminates invalid sub-intervals. The iteration step size refers to the proportion of each division, which is the golden ratio.

[0057] The preset granularity (taken as 100Hz under low signal-to-noise ratio, derived from noise level measurement under low signal-to-noise ratio, where noise is high and the sub-interval needs to be larger); the bisection method refers to an iterative search algorithm that continuously divides the search interval into two equal parts and eliminates invalid sub-intervals. The iteration step size refers to the proportion of each division, which is the division proportion of the bisection method;

[0058] The channel coherence window refers to a time window with the channel coherence time as its length. The maximum allowable frequency offset change refers to the maximum possible change in frequency offset within this window. The convergence threshold (500Hz in this scheme, derived from channel coherence time measurement; the change in frequency offset within the channel coherence time will not exceed this value) is a threshold used to determine whether the search has converged. When the difference in frequency offset estimates between two adjacent iterations is less than this threshold, the search is considered to have converged. The coarse frequency offset estimate refers to the preliminary frequency offset estimate obtained from the interval search. The signal-to-noise ratio (SNR) threshold of this module is obtained based on the performance intersection of the golden section method and the bisection method in high-speed scenarios. When the SNR is higher than the preset value, the convergence speed of the golden section method is better than that of the bisection method. When the SNR is lower than the preset value, the noise resistance of the bisection method is better than that of the golden section method.

[0059] The sub-interval granularity is obtained based on the noise level corresponding to the threshold. Under high signal-to-noise ratio, the noise is small, so the sub-interval can be smaller; under low signal-to-noise ratio, the noise is large, so the sub-interval needs to be larger. The calculation steps for convergence determination are as follows: take the channel coherence time Tc as the window, and the maximum allowable frequency offset change within the window is a preset value. When the frequency offset estimation difference between two adjacent iterations is less than the threshold, the search is determined to be converged, where Tc is the channel coherence time.

[0060] As an optional embodiment, the frequency offset fuzzy decoupling module is used to first perform differential detection on the frequency offset coarse estimation results of a consecutive preset number of frames. When the magnitude of the differential result is greater than half of the sampling frequency, it is determined that there is frequency offset folding. The corresponding folding factor is the rounded result of the differential result divided by the sampling frequency. Then, the coarse estimation result is corrected according to the detected folding factor.

[0061] The number of consecutive preset frames refers to a preset number of data frames that are consecutive in time (in this scheme, the number of preset frames is 5, which is derived from the adaptation calculation of channel coherence time and frame interval. The time of 5 frames is less than the coherence time, ensuring the validity of the differential result). This number of frames is determined by both the channel coherence time and the frame interval, ensuring the validity of the differential result.

[0062] Differential detection refers to a detection method that calculates the difference between the frequency offset results of adjacent frames; sampling frequency refers to the sampling frequency of the baseband signal; frequency offset folding refers to the phenomenon that when the Doppler frequency offset exceeds half of the baseband sampling frequency, the frequency offset will be folded into the sampling frequency range; folding factor refers to the integer coefficient used to correct the folded frequency offset.

[0063] The correction method is to equal the corrected frequency offset to the coarse estimate plus the folding factor multiplied by the sampling frequency, thus eliminating frequency offset folding ambiguity under high-speed movement. The continuous preset frame number differential detection in this module is based on the channel coherence time. If the preset frame number is less than the coherence time, the differential result is valid. This frame number is determined by both the channel coherence time and the frame interval. The folding factor detection steps are as follows:

[0064] The frequency offset coarse estimation results for a consecutive preset number of frames are differentially calculated. The differential result Δfi is equal to fi minus f{i-1}. When the magnitude of Δfi is greater than half of the sampling frequency fs, it is determined that frequency offset folding exists. The corresponding folding factor k is equal to the rounded result of Δfi divided by fs, where fi is the frequency offset coarse estimation result of the i-th frame, f{i-1} is the frequency offset coarse estimation result of the (i-1)-th frame, and fs is the baseband sampling frequency. The correction step is: the corrected frequency offset fcorrected is equal to the coarse estimation result frough plus k multiplied by fs.

[0065] As an optional embodiment, the frequency offset fine calibration module is used to first perform singular value decomposition on the covariance matrix composed of the decoupled frequency offset results, project the decomposed results onto the signal subspace corresponding to the first K largest singular values, remove the interference components of the noise subspace, then perform differential detection on the decoupled frequency offset results for a consecutive preset number of frames, use a preset value as a preset jump threshold, when the differential result is greater than the threshold, mark the frame as a jump frame, and finally generate a frequency offset confidence weight ranging from 0 to 1 based on the channel correlation coefficient and the jump mark.

[0066] The covariance matrix is ​​a matrix used to describe the statistical correlation of frequency offset results. Singular value decomposition is a matrix decomposition method that decomposes a matrix into a left singular matrix, a singular value matrix, and a right singular matrix, and is used to separate signal and noise components.

[0067] The signal subspace refers to the subspace composed of eigenvectors corresponding to large singular values, used to represent signal components. The noise subspace refers to the subspace composed of eigenvectors corresponding to small singular values, used to represent noise components. Differential detection refers to the detection method that calculates the difference between the frequency offset results of adjacent frames. A preset jump threshold is used (in this scheme, the value is 300Hz, derived from the calculation of the intersection of the differential results of frequency offset folding and normal jump in high-speed moving scenes. Normal frequency offset jump will not exceed this value, while the differential result of folding will reach the sampling frequency).

[0068] A jump frame refers to a frame where the frequency offset changes abruptly; the confidence weight is a weight used to represent the credibility of the frequency offset result. The calculation method is that the confidence is equal to the channel correlation coefficient multiplied by a set coefficient minus the set coefficient multiplied by the jump frame flag.

[0069] The jump threshold of this module is obtained based on the intersection of the differential result of frequency offset folding and the differential result of normal jump in high-speed moving scenarios. Normal frequency offset jump will not exceed the preset value, while the differential result of folding will reach the size of the sampling frequency.

[0070] The calculation steps for singular value decomposition are as follows: Perform SVD decomposition on the frequency offset covariance matrix R. The decomposition result R is equal to the conjugate transpose of U multiplied by Σ multiplied by V. Project the result onto the signal subspace corresponding to the first K largest singular values ​​and remove the components of the noise subspace. Here, R is the frequency offset covariance matrix, U is the left singular matrix, Σ is the singular value matrix, and V is the right singular matrix. The calculation steps for skip frame marking are as follows: Perform difference on the frequency offset results of multiple consecutive frames. When the difference result is greater than a preset value, mark the frame as a skip frame.

[0071] The confidence level is calculated as follows: the confidence level w is equal to the channel correlation coefficient r multiplied by the set coefficient minus the set coefficient multiplied by the jump frame flag isjump, where r is the channel correlation coefficient, isjump is the jump frame flag, and the confidence level ranges from 0 to 1.

[0072] As an optional embodiment, it also includes a multi-antenna federated fusion and privacy protection module, which is used to first perform scale normalization on the frequency offset characteristics of each antenna to eliminate the feature deviation of heterogeneous channels;

[0073] Then, the fusion weights are allocated according to the frequency offset confidence weights. The antenna weights with confidence values ​​lower than the preset value are reset to 0. Finally, the local features are distilled into global features of preset dimensions through a two-layer fully connected network. At the same time, Laplacian noise is added to the local features. The noise scale is the feature sensitivity divided by the privacy budget to protect the local privacy of multiple antennas.

[0074] Scale normalization refers to a processing method that converts the frequency offset characteristics of different antennas to a uniform scale. Heterogeneous channels refer to the phenomenon that the channel characteristics of different antennas differ.

[0075] Fusion weight refers to the weight used for multi-antenna feature fusion, with a preset threshold (in this scheme, the value is 0.3, which is derived from the confidence distribution calculation output by the fine calibration module; features below this threshold are considered noise).

[0076] A fully connected network refers to a neural network with all neurons connected. Knowledge distillation is a processing method that distills local features into global features without transmitting the original data. Laplace noise refers to random noise that follows a Laplace distribution, which can minimize the impact on feature accuracy while protecting privacy. Feature sensitivity refers to the maximum change in a feature. Privacy budget refers to the privacy parameter of differential privacy, which is used to control the degree of privacy protection.

[0077] The confidence threshold of this module is obtained based on the confidence distribution output by the calibration module; features below this threshold are considered noise. The normalization calculation steps are as follows: the normalized feature fnorm equals f minus the mean of f, divided by the standard deviation of f, and then multiplied by the antenna channel gain g, where f is the original frequency offset feature and g is the antenna channel gain. The fusion weight allocation steps are as follows:

[0078] The antenna weights with confidence levels below the threshold are reset to 0. The processing steps of the distillation network are as follows: a two-layer fully connected network distills the local features into global features of a preset dimension (in this scheme, the preset dimension is 128 dimensions, which is derived from the measurement of feature distillation accuracy and computing power adaptation). The calculation steps of privacy noise addition are as follows: Laplacian noise is added to the local features, and the noise scale is equal to Δf divided by ε, where Δf is the sensitivity of the feature and ε is the privacy budget.

[0079] As an optional embodiment, the closed-loop parameter control module is used to first perform differential calculation on the global frequency offset estimate to obtain the frequency offset jump rate, and use a preset value and a preset value as the jump rate threshold. When the jump rate is greater than the preset value, the search interval width is adjusted to the preset value and the search layer is increased to the preset layer. When the jump rate is less than or equal to the preset value, the search interval width is adjusted to the preset value and the search layer is reduced to the preset layer. The parameters corresponding to the intermediate rate are linearly interpolated and adjusted to form a closed loop with the interval search module.

[0080] The global frequency offset estimate refers to the final frequency offset estimate after multi-antenna fusion; differential calculation refers to the calculation method of the derivative of frequency offset with respect to time; and frequency offset hopping rate refers to the amount of frequency offset change per unit time.

[0081] These two hopping rate thresholds (in this scheme, they are set to 500Hz / s and 100Hz / s respectively, derived from the statistical distribution of terminal movement speed in high-speed mobile scenarios. When the terminal speed exceeds 1000km / h, the hopping rate will exceed 500Hz / s, and when the terminal speed is below 200km / h, the hopping rate will be below 100Hz / s) are obtained based on the statistical distribution of terminal movement speed in high-speed mobile scenarios.

[0082] Interval width refers to the frequency range of the search interval; search layer number refers to the number of iterations for interval shrinkage; linear interpolation refers to the interpolation method that calculates the value of the intermediate point using the values ​​of two known points; closed loop refers to the output of the control module being used as the input of the search module to achieve adaptive adjustment of parameters.

[0083] The jump rate threshold of this module is obtained based on the statistical distribution of the terminal's moving speed in high-speed mobile scenarios. The calculation steps of the jump rate are: the jump rate vf is equal to the derivative of the frequency offset f with respect to time t, that is, the change in frequency offset per unit time.

[0084] The calculation steps for parameter adjustment are as follows: when vf is greater than the preset value, the interval width is set to the preset value (1000Hz in this scheme, which is derived from the search accuracy adaptation calculation under high frequency bias jump rate), and the number of search layers is set to the preset number of layers (5 layers in this scheme, which is derived from the search efficiency and accuracy adaptation calculation).

[0085] When vf is less than or equal to the preset value, the interval width is set to the preset value (3000Hz in this scheme, derived from the search range adaptation calculation under low frequency offset rate), and the number of search layers is set to the preset number of layers (2 layers in this scheme, derived from the search efficiency and accuracy adaptation calculation); the linear interpolation steps for the intermediate value are as follows: the interval width w is equal to the preset value plus the preset value minus the preset value multiplied by the preset value minus vf divided by the preset value minus the preset value, and the number of search layers l is equal to the preset number of layers plus the preset number of layers minus the preset number of layers multiplied by the preset value minus vf divided by the preset value minus the preset value.

[0086] As an optional embodiment, it also includes an edge-end lightweight inference module. The edge-end lightweight inference module is used to first perform adaptive low-bit quantization on the interval search model, quantize the parameters of FP32 to the precision of INT8, and control the quantization error within a reasonable range obtained based on calibration data. Then, it eliminates low-priority search paths with confidence levels lower than preset values ​​based on frequency offset confidence weights. Finally, it adaptively adjusts the inference precision based on the computing power level of the edge terminal to adapt to terminals with different computing power.

[0087] Adaptive low-bit quantization refers to a method of quantizing high-precision parameters of a model into low-precision parameters to reduce the size and computational cost of the model; FP32 refers to 32-bit floating-point precision, and INT8 refers to 8-bit integer precision; preset threshold (in this solution, the value is 0.2, derived from the confidence distribution calculation output by the fine calibration module; paths below this threshold are invalid paths).

[0088] Search path pruning refers to eliminating invalid search paths to reduce computational load; computational power adaptation refers to adjusting inference accuracy based on the terminal's computational power to adapt to different terminals; the confidence threshold of this module is obtained based on the confidence distribution output by the calibration module, and paths below this threshold are invalid paths; the calculation steps for adaptive quantization are as follows: the model parameters are calibrated based on channel data of the previous preset number of frames (10 frames in this scheme, derived from calibration accuracy and computational load adaptation calculations), and the FP32 parameters are quantized to INT8 accuracy; the calculation steps for search path pruning are as follows:

[0089] Low-priority paths with confidence levels below the threshold are eliminated; the processing steps for computing power adaptation are as follows: based on the computing power level of the edge terminal, the inference accuracy is adaptively adjusted to adapt to terminals with different computing power.

[0090] As an optional embodiment, it also includes an anomaly calibration module, which is used to first use a sliding window of a preset length to perform a Grubbs statistical test on the global frequency offset estimate within the window, and remove abnormal data that deviate from the confidence interval of the mean plus or minus a preset multiple of the standard deviation. Then, based on the effective frequency offset data after removing anomalies, a mini-batch stochastic gradient descent method is used with a preset value as the learning rate to incrementally update the calibration parameters of the initial search interval.

[0091] A sliding window is a window that slides across time to perform statistical processing on continuous frame data. The window length is a preset value (8 in this scheme, derived from channel coherence time measurement; the window length is less than the coherence time to ensure the validity of the statistical results). It is based on the channel coherence time, and the statistical results are valid because the window length is less than the coherence time. The Grubbs statistic test is a statistical test method used to identify outliers. By calculating the Grubbs statistic for each data point and comparing it with the critical value, outlier data that deviates from the confidence interval can be identified.

[0092] The mean refers to the average value of the data within the window, the standard deviation refers to the standard deviation of the data within the window, the preset multiple (in this scheme, the value is 3 times, derived from the confidence interval calculation of normal data distribution, which includes most normal data) and the standard deviation confidence interval refer to the confidence interval that includes most normal data.

[0093] Mini-batch stochastic gradient descent is an incremental optimization algorithm that uses a small batch of valid data to perform gradient descent updates each time. This reduces computational cost while maintaining update stability. The learning rate refers to the step size of the gradient descent, which controls the update speed. Incremental update refers to the method of gradually updating parameters, which is used to correct long-term parameter drift.

[0094] The sliding window length of this module is obtained based on the channel coherence time. The statistical results are valid if the window length is less than the coherence time. This window length is determined by both the channel coherence time and the frame interval. The learning rate (0.01 in this scheme, derived from the incremental update convergence speed calculation, which can ensure the stability of incremental updates) is obtained based on the incremental update convergence speed. This learning rate can ensure the stability of incremental updates.

[0095] The calculation steps for anomaly detection are as follows: statistical analysis is performed on the frequency offset data within the window, the mean μ and standard deviation σ of the window are calculated, and data that deviates from μ plus or minus a preset multiple σ are identified as anomaly data; the calculation steps for incremental update are as follows: the mini-batch stochastic gradient descent method is used, the learning rate is set to a preset value, and the calibration parameters of the initial search interval are incrementally updated.

[0096] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A range-search Doppler frequency offset estimation system for high-speed mobile scenarios, characterized in that, include: The channel sensing module is used to receive fast time-varying channel parameters and terminal motion parameters, perform autocorrelation calculation on the received channel impulse response to obtain the channel coherence time threshold, calculate the maximum theoretical frequency offset at the current motion speed according to the Doppler formula, eliminate invalid frequency offset intervals that exceed the range, and output the initial frequency offset search interval. An adaptive multi-level interval shrinkage search module is used to receive the initial search interval and closed-loop control parameters. First, it detects the signal-to-noise ratio of the current channel, and then adaptively adjusts the division granularity and shrinkage algorithm of the sub-interval based on the signal-to-noise ratio. Finally, it uses the maximum allowable frequency offset change within the channel coherence window as the convergence threshold to determine search convergence and outputs a coarse frequency offset estimation result. The frequency offset fuzzy decoupling module is used to receive the frequency offset coarse estimation result, perform differential detection on the frequency offset coarse estimation result for a consecutive preset number of frames, identify the folding factor of the sampling frequency offset as an integer multiple, correct the coarse estimation result according to the detected folding factor, and output the decoupled frequency offset result. The frequency offset fine-tuning module is used to receive the decoupled frequency offset result, perform singular value decomposition on the decoupled frequency offset result, project the result onto the signal subspace to remove noise components, perform differential detection to mark the jump frame on the decoupled frequency offset result for a consecutive preset number of frames, generate frequency offset confidence weight based on the channel correlation coefficient and the jump mark, and output the frequency offset confidence weight. The closed-loop parameter control module receives the global frequency offset estimate, performs differential calculation on the global frequency offset estimate to obtain the frequency offset jump rate, adaptively adjusts the search interval width and search layer number according to the jump rate, performs linear interpolation adjustment on the parameters corresponding to the intermediate rate, and outputs the control parameters for the interval search, forming a closed loop with the adaptive multi-level interval shrinkage search module.

2. The interval-search Doppler frequency offset estimation system for high-speed moving scenarios according to claim 1, characterized in that, It also includes a space-time-frequency preprocessing module, which is used to perform space-time joint filtering on the received baseband signal. It adopts an adaptive filtering window of a preset size. The weight of the filter is adaptively adjusted according to the delay spread of the current channel to separate short-delay multipath coupling components. Then, the filtered signal is cross-correlated with the local synchronization preamble symbol to complete symbol-level alignment and control the alignment error within a set range. Finally, frequency domain noise whitening processing is performed.

3. The interval-search Doppler frequency offset estimation system for high-speed mobile scenarios according to claim 1, characterized in that, The channel sensing module is used to first perform autocorrelation calculation on the received channel impulse response to obtain the channel coherence time threshold. Then, through the inertial measurement unit built into the terminal, it collects the terminal's moving speed, three-axis acceleration and heading angle parameters at a preset sampling rate. Finally, it calculates the maximum theoretical frequency offset at the current moving speed according to the Doppler formula, eliminates invalid frequency offset intervals that exceed the range, and determines the initial search interval.

4. The interval-search Doppler frequency offset estimation system for high-speed moving scenarios according to claim 1, characterized in that, It also includes an interval search module, which first detects the signal-to-noise ratio (SNR) of the current channel and sets a preset threshold. When the SNR is higher than the preset threshold, the initial search interval is divided into sub-intervals with a preset granularity. The golden section method is used to shrink the sub-intervals, and the iteration step size is set to a preset value. When the SNR is not higher than the preset threshold, the initial search interval is divided into sub-intervals with a preset granularity and the bisection method is used to shrink the sub-intervals, with the iteration step size set to a preset value. Finally, the maximum allowable frequency offset change within the channel coherence window is used as the convergence threshold to determine search convergence and output the coarse frequency offset estimation result.

5. The interval-search Doppler frequency offset estimation system for high-speed moving scenarios according to claim 4, characterized in that, The frequency offset fuzzy decoupling module is used to first perform differential detection on the frequency offset coarse estimation result of a consecutive preset number of frames. When the magnitude of the differential result is greater than half of the sampling frequency, it is determined that there is frequency offset folding. The corresponding folding factor is the rounded result of the differential result divided by the sampling frequency. Then, the coarse estimation result is corrected according to the detected folding factor.

6. The interval-search Doppler frequency offset estimation system for high-speed moving scenarios according to claim 5, characterized in that, The frequency offset fine-tuning module is used to first perform singular value decomposition on the covariance matrix composed of the decoupled frequency offset results, project the decomposed results onto the signal subspace corresponding to the first K largest singular values, remove the interference components of the noise subspace, then perform differential detection on the decoupled frequency offset results for a preset number of consecutive frames, and use a preset value as a preset jump threshold. When the differential result is greater than the threshold, the frame is marked as a jump frame. Finally, based on the channel correlation coefficient and the jump mark, a frequency offset confidence weight ranging from 0 to 1 is generated.

7. The interval-search Doppler frequency offset estimation system for high-speed mobile scenarios according to claim 6, characterized in that, It also includes a multi-antenna federated fusion and privacy protection module, which is used to first perform scale normalization on the frequency offset characteristics of each antenna to eliminate the feature deviation of heterogeneous channels; Then, the fusion weights are allocated according to the frequency offset confidence weights. The antenna weights with confidence values ​​lower than the preset value are reset to 0. Finally, the local features are distilled into global features of preset dimensions through a two-layer fully connected network. At the same time, Laplacian noise is added to the local features. The noise scale is the feature sensitivity divided by the privacy budget to protect the local privacy of multiple antennas.

8. The interval-search Doppler frequency offset estimation system for high-speed moving scenarios according to claim 7, characterized in that, The closed-loop parameter control module is used to first perform differential calculation on the global frequency offset estimate to obtain the frequency offset jump rate. Using a preset value and a preset value as the jump rate threshold, when the jump rate is greater than the preset value, the search interval width is adjusted to the preset value and the search layer is increased to the preset layer. When the jump rate is less than or equal to the preset value, the search interval width is adjusted to the preset value and the search layer is reduced to the preset layer. The parameters corresponding to the intermediate rate are linearly interpolated and adjusted to form a closed loop with the interval search module.

9. The interval-search Doppler frequency offset estimation system for high-speed moving scenarios according to claim 1, characterized in that, It also includes a lightweight edge inference module, which first performs adaptive low-bit quantization on the interval search model to quantize the FP32 parameters to INT8 precision, and controls the quantization error within a reasonable range based on calibration data. Then, it eliminates low-priority search paths with confidence levels lower than preset values ​​based on frequency offset confidence weights. Finally, it adaptively adjusts the inference precision based on the computing power level of the edge terminal to adapt to terminals with different computing power.

10. The interval-search Doppler frequency offset estimation system for high-speed moving scenarios according to claim 1, characterized in that, It also includes an anomaly calibration module, which first uses a sliding window of a preset length to perform a Grubbs statistical test on the global frequency offset estimate within the window, and removes outlier data that deviates from the confidence interval of the mean plus or minus a preset multiple of the standard deviation. Then, based on the effective frequency offset data after removing the anomalies, it uses a mini-batch stochastic gradient descent method with a preset learning rate to incrementally update the calibration parameters of the initial search interval.

Images