A method and system for estimating delay of Wi-Fi channel state information

Abstract

This invention discloses a method and system for estimating delay of Wi-Fi channel state information (CSI). The method includes: preprocessing collected multi-antenna CSI data to obtain a multi-antenna CSI matrix; constructing a joint Bayesian model for multiple antennas, extracting the principal path of each antenna and performing consistency detection, calculating the fused delay value, delay uncertainty, and channel anomaly identifier to form state features; employing a lightweight DQN combining depthwise separable convolution and fully connected layers, training the DQN through a multi-objective reward function, and outputting the optimal pilot configuration action; executing the optimal pilot configuration action and mapping it to pilot configuration parameters, generating and transmitting time-domain orthogonal multi-antenna pilot signals, and outputting a new CSI matrix; updating the network parameters of the multi-antenna Bayesian model and the lightweight DQN, and performing closed-loop optimization for dynamic channel adaptation. This invention effectively overcomes the impact of hardware errors on delay estimation by organically combining joint preprocessing of multi-antenna data with Bayesian estimation.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication, specifically Wi-Fi communication technology, and relates to a method and system for estimating the delay of Wi-Fi channel state information. Background Technology

[0002] Wi-Fi channel state information delay estimation refers to the technique of estimating the multipath propagation delay of Wi-Fi signals from the transmitter to the receiver by extracting the channel state information (CSI) formed during the transmission of Wi-Fi signals and combining it with signal processing algorithms. In a multi-antenna MIMO architecture, CSI data needs to be collected by multiple antennas in a coordinated manner to improve the stability of delay estimation by utilizing spatial diversity characteristics.

[0003] Currently, Wi-Fi channel state information (CSI) delay estimation technologies are mainly divided into two categories: one is delay estimation technology based on single-antenna CSI, which converts the frequency domain CSI into the delay domain through Fourier transform, or uses a simple Bayesian model to estimate multipath delay. Some schemes will separately calibrate hardware errors such as crystal oscillator offset and radio frequency phase noise. The other is resource control technology combined with intelligent algorithms, such as using traditional deep Q networks to make decisions on pilot configuration in order to balance delay estimation accuracy and pilot overhead. However, most of these technologies use fully connected network structures and do not include energy consumption in the decision-making objectives.

[0004] However, existing technologies have some problems. For example, multiple hardware errors coupled together lead to low estimation accuracy. Existing technologies mostly deal with crystal oscillator offset and radio frequency phase noise separately, without considering the coupling relationship between the phase inconsistency of multiple antennas and the first two types of errors. For example, after calibrating the crystal oscillator offset of a single antenna, there is still a problem of inconsistent phase references among multiple antennas due to hardware differences. This results in poor spatial consistency of CSI data from multiple antennas, which ultimately makes the uncertainty of time delay estimation often exceed 1ns, failing to meet the requirements of high-precision sensing. Summary of the Invention

[0005] Purpose of the invention: To address the problem of low estimation accuracy caused by the coupling of multiple hardware errors, this invention provides a method and system for estimating the delay of Wi-Fi channel state information. By organically combining multi-antenna data joint preprocessing with Bayesian estimation, the invention effectively overcomes the impact of hardware errors on delay estimation.

[0006] Technical Solution: To achieve the above objectives, this invention provides a method for estimating the delay of Wi-Fi channel state information, comprising the following steps:

[0007] S1: Preprocess the acquired multi-antenna CSI data to eliminate hardware errors such as crystal oscillator offset, radio frequency phase noise, and phase inconsistency of multiple antennas, and obtain the multi-antenna CSI matrix;

[0008] S2: Based on the multi-antenna CSI matrix, construct a multi-antenna joint Bayesian model, extract the main path of each antenna and perform consistency detection, calculate the fusion delay value, delay uncertainty and channel anomaly identifier to form state features;

[0009] S3: A lightweight DQN combining depthwise separable convolution and fully connected layers is used. The state features are used as input, and the DQN is trained through a multi-objective reward function. Convergence verification is performed, and the optimal pilot configuration action is output.

[0010] S4: Execute the optimal pilot configuration action and map it to pilot configuration parameters, generate and send time-domain orthogonal multi-antenna pilot signals, receive the feedback of the new CSI matrix, and output the new CSI matrix after verification;

[0011] S5: Repeat step S1 for preprocessing the new CSI matrix to obtain the preprocessed new CSI matrix. Update the network parameters of the multi-antenna Bayesian model and the lightweight DQN, and perform closed-loop optimization for dynamic channel adaptation.

[0012] Further, the elimination of crystal oscillator offset in step S1 includes: calculating the average phase difference between adjacent frames of a single antenna pair, fusing the phase differences of all antenna pairs to determine the crystal oscillator offset frequency, and eliminating the phase linear drift of the same subcarrier in adjacent frames in reverse based on the crystal oscillator offset frequency to obtain the true phase of the multi-antenna after eliminating the crystal oscillator offset.

[0013] The elimination of radio frequency phase noise includes: smoothing the phase noise of multiple antennas through Kalman filtering, using the weighted average of the phases of multiple antennas as the observation value, with the weight of the weighted average being the reciprocal of the phase variance of each antenna, and performing iterative prediction and correction to eliminate random phase noise introduced by the radio frequency link;

[0014] The elimination of hardware errors caused by phase inconsistency among multiple antennas includes: in the process of eliminating crystal oscillator offset and Kalman filtering smoothing, the phase inconsistency among multiple antennas is eliminated by fusing observation data from multiple antennas.

[0015] Furthermore, the construction of the multi-antenna joint Bayesian model in step S2 includes: constructing a multi-antenna joint Bayesian model based on the linear mapping relationship between frequency domain CSI and multipath delay. The multi-antenna joint Bayesian model includes multi-antenna prior constraints, which include sparse priors of multi-antenna main path amplitude deviation and multipath delay amplitude vector. The sparse priors are used to ensure that only the main path has a non-zero amplitude.

[0016] Furthermore, the calculation of the channel anomaly identifier in step S2 includes:

[0017] A1: Calculate the average signal-to-noise ratio (SNR) of all antennas and subcarriers, and set a dynamic threshold based on the SNR;

[0018] A2: Calculate the residual energy between CSI observations and predictions. If the residual energy exceeds the dynamic threshold, mark it as an anomaly.

[0019] A3: Estimate the Doppler frequency shift by the CSI correlation of adjacent frames, and determine the anomaly type by combining residual energy, time delay uncertainty and Doppler frequency shift.

[0020] Furthermore, in step S3, the multi-objective reward function is a weighted sum of accuracy reward, overhead reward, quality reward and energy consumption reward, with a total weight of 1. Accuracy reward is negatively correlated with time delay uncertainty, overhead reward is negatively correlated with pilot density, quality reward is positively correlated with signal-to-noise ratio and is halved in case of anomalies, and energy consumption reward is negatively correlated with pilot energy consumption.

[0021] Furthermore, the lightweight DQN structural design in step S3 includes: replacing the existing fully connected network of DQN with a depthwise separable convolution + fully connected structure, and defining network layer parameters and hyperparameters;

[0022] The network structure includes:

[0023] Input layer: Fully connected input;

[0024] Feature extraction layer: The type is depthwise separable convolutional layer;

[0025] Hidden layer: Type is fully connected layer;

[0026] Output layer: The type is fully connected output.

[0027] Furthermore, the mapping of pilot configuration parameters in step S4 includes pilot density, subcarrier strategy, applicable channel scenario, and pilot energy consumption;

[0028] The generation of time-domain orthogonal multi-antenna pilot signals includes: generating pilot signals for each transmit antenna based on the Zadoff-Chu sequence, assigning different dedicated root indices to different transmit antennas, mapping the pilot signals of different transmit antennas to different time-domain segments of the frame, and the receiving end of the Wi-Fi system extracting the corresponding antenna pilot signals through a time-domain window.

[0029] Furthermore, the update of the multi-antenna Bayesian model in step S5 includes: re-estimating the noise variance of the model based on the preprocessed new CSI matrix samples, with the number of samples increasing with each feedback; if the amplitude deviation of the main diameter of the multi-antenna exceeds a set value, the prior covariance of the model is increased; otherwise, the prior covariance remains unchanged.

[0030] The update of network parameters in lightweight DQN includes: calculating new state features based on the preprocessed new CSI matrix, generating new experience by combining pilot configuration actions, total reward and next state, and storing it in the experience replay pool; randomly sampling experience from the experience replay pool, repeatedly training the network, updating the parameters of the main network and the target network, and verifying the accuracy of action selection in mobile device scenarios.

[0031] This invention provides a Wi-Fi channel state information delay estimation system, comprising:

[0032] The multi-antenna data preprocessing module is used to collect raw channel state information data from multiple antennas, eliminate hardware errors through joint calibration and filtering of multiple antennas, and output the preprocessed multi-antenna CSI matrix.

[0033] The state feature extraction module is used to receive the preprocessed multi-antenna CSI matrix, calculate the delay value, delay uncertainty and channel anomaly identifier through Bayesian estimation of multi-antenna fusion, and form state features;

[0034] The intelligent decision-making module receives state features, performs decision analysis through a lightweight deep Q-network, and outputs the optimal pilot configuration action.

[0035] The resource control and verification module is used to receive the optimal pilot configuration action, generate and send multi-antenna pilot signals, receive the new CSI matrix fed back by the terminal, and verify the control effect.

[0036] The closed-loop optimization module is used to receive the new CSI matrix and update the Bayesian model parameters in the state feature extraction module and the network parameters in the intelligent decision-making module.

[0037] Furthermore, the multi-antenna data preprocessing module includes a crystal oscillator offset calibration unit and a Kalman filter smoothing unit. The crystal oscillator offset calibration unit calculates the crystal oscillator offset frequency by fusing the phase difference of all antenna pairs and eliminates linear phase drift. The Kalman filter smoothing unit performs iterative prediction and correction by fusing multi-antenna observations to eliminate random phase noise.

[0038] The state feature extraction module includes a Bayesian delay estimation unit and a dynamic anomaly identification unit. The Bayesian delay estimation unit constructs a joint Bayesian model containing principal path consistency constraints and calculates the fused delay value. The dynamic anomaly identification unit adaptively adjusts the judgment threshold based on the signal-to-noise ratio and generates anomaly types.

[0039] The intelligent decision-making module includes a lightweight DQN structure and a multi-objective reward function unit. The lightweight DQN structure uses a combination of depthwise separable convolution and fully connected layers. The multi-objective reward function unit includes a weighted combination of accuracy reward, overhead reward, quality reward and energy reward.

[0040] The resource control and verification module includes an action-parameter mapping unit and a time-domain orthogonal pilot generation unit. The action-parameter mapping unit converts discrete actions into pilot density and subcarrier policy parameters, while the time-domain orthogonal pilot generation unit generates orthogonal pilot signals for different antennas based on the Zadoff-Chu sequence.

[0041] The closed-loop optimization module includes a Bayesian model update unit and a lightweight DQN update unit. The Bayesian model update unit updates the noise variance and prior covariance based on accumulated samples, while the lightweight DQN update unit periodically retrains the network parameters through empirical replay.

[0042] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0043] 1. This invention effectively overcomes the impact of hardware errors on time delay estimation by organically combining multi-antenna data joint preprocessing with Bayesian estimation. Specifically, multi-antenna joint crystal oscillator offset calibration eliminates linear phase drift, multi-antenna phase noise Kalman filtering smooths and suppresses random phase fluctuations, and multi-antenna fused Bayesian estimation improves the reliability of time delay estimation through principal path consistency constraints, thus achieving high-precision time delay estimation results even in complex real-world hardware environments.

[0044] 2. The method of this invention achieves adaptive optimization of pilot configuration by introducing a lightweight deep Q-network intelligent decision-making core. Based on multi-dimensional state characteristics such as latency uncertainty and anomaly identification, this network learns a resource control strategy that achieves the optimal balance between accuracy, overhead, and energy consumption through the guidance of a multi-objective reward function. This solves the problem that static configuration strategies cannot adapt to dynamic scenarios, and significantly improves the system's adaptability and energy efficiency in complex environments.

[0045] 3. This invention constructs a complete closed-loop system of perception-decision-control-feedback. By continuously updating the Bayesian model and deep Q-network parameters, the system can track long-term changes in channel characteristics and continuously optimize latency estimation performance. This closed-loop optimization mechanism ensures the effectiveness and stability of the method in long-term deployment, providing a guarantee for the reliable operation of Wi-Fi latency estimation technology in practical applications. Attached Figure Description

[0046] Figure 1 This is a flowchart of the method of the present invention;

[0047] Figure 2 This is a system architecture diagram of the system of the present invention. Detailed Implementation

[0048] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0049] Example 1:

[0050] like Figure 1 As shown, this embodiment provides a method for estimating the delay of Wi-Fi channel state information, including the following steps:

[0051] Step 1: Acquisition and preprocessing of multi-antenna CSI data: Collect raw channel state information (CSI) data of the multi-antenna Wi-Fi system, eliminate hardware errors such as crystal oscillator offset, radio frequency phase noise and phase inconsistency of multiple antennas, and obtain a clean multi-antenna CSI matrix after preprocessing.

[0052] The elimination of crystal oscillator offset includes: calculating the average phase difference between adjacent frames of a single antenna pair, fusing the phase differences of all antenna pairs to determine the crystal oscillator offset frequency, and eliminating the phase linear drift of the same subcarrier in adjacent frames in reverse based on the crystal oscillator offset frequency to obtain the true phase of the multi-antenna after eliminating the crystal oscillator offset.

[0053] The elimination of radio frequency phase noise includes: smoothing the phase noise of multiple antennas through Kalman filtering, using the weighted average of the phases of multiple antennas as the observation value, with the weight of the weighted average being the reciprocal of the phase variance of each antenna, and performing iterative prediction and correction to eliminate random phase noise introduced by the radio frequency link;

[0054] In the process of eliminating crystal oscillator offset and smoothing with Kalman filtering, phase inconsistencies between multiple antennas are eliminated by fusing observation data from multiple antennas.

[0055] A Wi-Fi system consists of a receiver and a transmitter. The receiver is the hardware device responsible for receiving Wi-Fi signals, demodulating the signals, and generating CSI feedback. The transmitter is the hardware device responsible for transmitting Wi-Fi signals; its core function is to generate and output signals, providing a signal source for CSI data acquisition and delay estimation. CSI is a core parameter in a Wi-Fi system that describes the influence of the channel on the signal during its propagation from the transmitter to the receiver. Wi-Fi uses OFDM (Orthogonal Frequency Division Multiplexing) technology, dividing the entire Wi-Fi bandwidth into multiple mutually orthogonal subcarriers. Each subcarrier independently carries a signal and is independently affected by the channel during propagation. By measuring the CSI of each subcarrier, the state of the entire Wi-Fi channel can be fully reconstructed.

[0056] Based on the above Wi-Fi system, the implementation of step 1 in this embodiment specifically includes the following sub-steps:

[0057] Step 1.1: Acquisition of raw CSI data with multiple antennas: Based on the IEEE 802.11ax protocol and using a 2×2 MIMO architecture with 2 transmit antennas + 2 receive antennas, raw CSI data is acquired through an Intel 5300 network card and CSITool, covering the 2.4GHz band in low-bandwidth scenarios and the 5GHz band in high-bandwidth scenarios, obtaining subcarrier-level CSI information of continuous frames, including subcarrier amplitude and phase information.

[0058] Multi-antenna CSI matrix: The original CSI data exists in the form of a complex channel coefficient matrix, denoted as... Where: C represents the complex field; The effective number of subcarriers, such as under a 20MHz bandwidth. ; The number of frames to be continuously captured, such as 100 frames; 2 represents the number of transmitting antennas; Number of receiving antennas; original multi-antenna CSI matrix These are all the original single-antenna CSI complex coefficients. A 4D matrix organized by dimension The element in the x-th row, rx-th column, i-th sub-dimension, and t-th time dimension is... .

[0059] By modifying the Wi-Fi network card driver, such as the Linux driver for the Intel 5300 network card, or using open-source tools such as CSI Tool, the complex coefficients of each subcarrier in each frame can be extracted from the physical layer. Where i is the subcarrier index, t is the frame index, tx is the transmit antenna index, and rx is the receive antenna index. The original CSI amplitude is given by e, where e is the natural constant and j is the imaginary unit. The original CSI phase is measured directly from the physical layer chip of the Wi-Fi network card.

[0060] Hardware errors during the acquisition process include crystal oscillator offset (SFO), radio frequency phase noise, and phase inconsistency among multiple antennas.

[0061] Step 1.2: Multi-antenna joint SFO calibration: Eliminate crystal oscillator offset differences between multiple antennas, reduce single antenna noise interference, and output the true phase of multiple antennas with SFO drift eliminated.

[0062] Average phase difference of a single antenna: The average phase difference between frame t and frame t-1 of a single antenna (rx, tx) is calculated using the following expression:

[0063]

[0064] in, The average phase difference, The original CSI phase of frame t; The original CSI phase is for frame t-1.

[0065] Multi-antenna SFO calibration: The phase difference of all antenna pairs is fused, and the crystal oscillator offset frequency is calculated. The expression is:

[0066]

[0067] in, This is the crystal oscillator offset frequency, used to quantify the interference of crystal oscillator offset on phase. Typical value is ±50Hz, and the calibration target is within ±1Hz. The frame interval is determined by the Wi-Fi protocol, such as the 802.11ax protocol. .

[0068] Phase calculation after calibration: Crystal oscillator offset frequency This will cause the phase of the same subcarrier in adjacent frames to increase linearly, inversely eliminating the phase drift caused by SFO and obtaining the true phase, expressed as:

[0069]

[0070] in, The true phase after SFO elimination; t is the frame index.

[0071] By introducing multi-antenna summation, the interference of single-antenna noise on SFO fitting is reduced, the calibration accuracy is improved, and the processed output is the true phase of the multi-antenna system that eliminates SFO drift. , which serves as the input for multi-antenna phase noise Kalman filtering.

[0072] Step 1.3: Multi-antenna phase noise Kalman filtering smoothing: The nonlinearity of the RF link introduces random phase noise. Kalman filtering (KF) is used to smooth the phase and eliminate random phase noise. The stability of the filter is improved by fusing multi-antenna observations, and a clean multi-antenna CSI matrix H is output.

[0073] Based on the Kalman-Frankel formula, multi-antenna observation fusion is added, and the weighted average of the phases of multiple antennas is taken as the Kalman-Frankel observation value, with the weight being the reciprocal of the variance of the phase of each antenna. Phase update and noise cancellation are achieved through Kalman filtering. The core is to first predict the current phase based on the historical phase, and then correct the prediction with the fused observations from multiple antennas, iteratively reducing random noise.

[0074] In this embodiment, the tx,rx antenna is used as an example. The specific process is as follows:

[0075] Phase prediction: For each antenna pair (tx, rx) and subcarrier i, the observed value is Based on the smoothed phase of the previous frame, i.e., the posterior phase Predict the phase of the current frame The state equation is:

[0076]

[0077] in, This is a priori estimate of the phase of the t-th frame; This is the posterior estimate of the phase of the (t-1)th frame; The variance Q is the process noise, determined by the hardware manual.

[0078] Multi-antenna observation fusion: The phase noise of different antennas is random. Using the reciprocal of the phase variance of each antenna as a weight, the phase observations of all antennas are fused by weighted averaging to obtain a more stable fused observation. The expression is:

[0079]

[0080] in, These are multi-antenna fused observations; The phase variance of the antenna at the tx,rx th frame is estimated using historical CSI data from 100 frames.

[0081] By establishing the error relationship between prediction and observation through observation equations, it is clear that the reason for inaccurate prediction is process noise. and observation noise v t The expression for the observation equation is:

[0082]

[0083] in, R represents the observation noise, and R is the observation noise variance, which is measured using Wi-Fi idle subcarriers.

[0084] The correction weights for the observations are determined by Kalman gain calculation:

[0085]

[0086] Among them, K t Kalman gain; The prior covariance reflects the uncertainty of the prior predicted phase; the initial value is... =10 -3 R is the variance of observation noise; K t Controlling the weights of observations on updates, if prior covariance If K is small, then tIf the variance of the observation noise R is small, then K... t Larger scales require more observation for correction.

[0087] State update: Correct the prior predicted phase with fused observations to obtain the smoothed final phase. The expression is:

[0088]

[0089] in, The posterior estimate of the phase for frame t is the smoothed final phase; the error term... Multiplied by Kalman gain K t The post-correction of the prior phase is equivalent to eliminating noise according to the reliability ratio. Random radio frequency phase noise will be gradually canceled out by this prediction-correction iteration.

[0090] Posterior covariance update: The posterior covariance reflects the uncertainty of the updated phase. With each iteration, the uncertainty decreases, and the noise reduction effect is enhanced. The expression is:

[0091]

[0092] in, For posterior covariance; Let be the a priori covariance.

[0093] Radio frequency phase noise is a random, high-frequency fluctuating error, while the phase change of the real channel is slow and continuous. Kalman filtering gradually removes random noise and preserves the true phase through an iterative process of historical phase prediction and multi-antenna observation correction, ultimately achieving phase smoothing and outputting a noise-free, clean multi-antenna CSI matrix H.

[0094] Step 2: High-precision sensing front-end of multi-antenna fusion: Based on the preprocessed and clean multi-antenna CSI matrix H, a multi-antenna joint Bayesian model is constructed, the principal path of each antenna is extracted and consistency detection is performed, the fusion delay value, delay uncertainty and channel anomaly identifier are calculated, and the state feature S is formed.

[0095] The multi-antenna joint Bayesian model is constructed based on the linear mapping relationship between frequency domain CSI and multipath delay. The multi-antenna joint Bayesian model includes multi-antenna prior constraints, which include the multi-antenna main path amplitude deviation being less than 10% and the sparse prior of the multipath delay amplitude vector. The sparse prior guarantees that only the main path has a non-zero amplitude.

[0096] Extracting the main diameter of each antenna and performing consistency detection includes: taking the delay grid point with the largest posterior mean amplitude of a single antenna pair as the main diameter index of that antenna pair; if the difference between the main diameter indices of all antenna pairs does not exceed 1, then the main diameters are determined to be consistent and any index is taken as the final main diameter index; if the difference exceeds 1, then the main diameter indices of most antenna pairs are taken as the final main diameter index.

[0097] The calculation of channel anomaly identification includes: first, calculating the average signal-to-noise ratio (SNR) of all antennas and subcarriers, and setting a dynamic threshold based on the SNR; calculating the residual energy between the CSI observation and prediction values, and marking an anomaly if the residual energy exceeds the dynamic threshold; at the same time, estimating the Doppler shift through the correlation of CSI between adjacent frames, and determining the anomaly type by combining the residual energy, time delay uncertainty and Doppler shift.

[0098] Based on the above, the implementation of step S2 in this embodiment specifically includes the following sub-steps:

[0099] Step 2.1: Construct a multi-antenna Bayesian delay estimation model: Wi-Fi signals propagate through multiple paths, and the CSI frequency domain response satisfies a linear model. Introduce a multi-antenna main path consistency constraint, with the main path amplitude deviation <10%, and construct a multi-antenna joint Bayesian model to quantify the uncertainty of delay estimation.

[0100] The multi-antenna linear model maps the frequency domain CSI to the time delay domain, providing a basic model for Bayesian estimation. Its expression is:

[0101]

[0102] in, A is a clean CSI vector for a single antenna pair, derived from a clean multi-antenna CSI matrix H; A is a pilot matrix with elements... , Let i be the frequency of the i-th subcarrier. The delay is the time delay of the l-th delay grid point; It is a multipath delay amplitude vector, with sparsity: only the principal path has non-zero values; This is the residual noise vector. Let I be the noise variance, and I be the identity matrix.

[0103] The prior constraints for multiple antennas are expressed as follows:

[0104]

[0105] in, Let be the prior covariance matrix; α1~Gamma(0.1,0.1) is the sparse prior, ensuring that only the principal diameter has a significant amplitude;

[0106] Constraint: The amplitude deviation of the main diameter of the multi-antenna is <10%, i.e. , It serves as the primary path index. Multipath interference is reduced through sparse priors, and spatial consistency across multiple antennas is ensured through primary path constraints.

[0107] Posterior Distribution of Multiple Antennas: Combining the linear model of multiple antennas and the prior distribution of multipath amplitudes, the posterior distribution of multiple antennas is derived, thus completing the construction of the Bayesian model; based on Bayes' theorem, the product of the posterior distributions of each antenna is calculated, i.e., the joint posterior distribution of multiple antennas, expressed as:

[0108]

[0109] in, A set of multipath amplitude vectors The posterior probability distribution, For noise variance; This is the posterior mean of a single antenna, i.e., the estimated multipath amplitude; This represents the a posteriori covariance of a single antenna, reflecting the uncertainty in the estimation.

[0110] The processed output includes the multi-antenna posterior results, including the multi-antenna posterior mean. With covariance .

[0111] The main path consistency constraint will be encoded into For example, by adjusting the covariance of the corresponding dimensions of the principal diameter for different antenna pairs, the principal diameter amplitude distributions of each antenna pair are made correlated, such as the covariance of the principal diameter dimension being non-zero. This ensures that in subsequent posterior inference, the principal diameter amplitude naturally satisfies the constraint of deviation <10%. In Bayesian inference, [the following is incomplete and requires further context:] Prior information about the relevance of the encoding directly modifies the parameters of the posterior distribution without changing its functional form. This design preserves the simplicity of posterior computation, as the product form is easy to solve, while prior constraints ensure spatial consistency across multiple antennas, conforming to the logic of prior guiding posterior in Bayesian inference.

[0112] Step 2.2: Calculation of delay and uncertainty of multi-antenna fusion: Based on the posterior results of multi-antenna fusion, extract the main path index of each antenna, determine the final main path through consistency detection, and calculate the fusion delay value and uncertainty.

[0113] Single-antenna main path index extraction: Locate the main path in a multipath propagation to avoid multipath interference. The main path is the time delay grid point with the largest posterior mean amplitude, expressed as:

[0114]

[0115] in, For the main diameter index of a single antenna pair; posterior mean The l-th element; L=200 is the number of delay grid points.

[0116] Multi-antenna main diameter consistency detection: If the main diameter indices of all antenna pairs differ... If the main paths are consistent, it is considered a scenario where the main path is clearly defined, and the main path is taken. Otherwise, take the major diameter index of the majority antenna. To avoid misjudgment of the main diameter due to noise from a single antenna and improve the reliability of the main diameter.

[0117] Final delay calculation: The actual delay corresponding to the delay grid point is: , To achieve delay resolution, with B=320MHz as the total bandwidth, the delay values ​​of multiple antennas are fused to calculate the final delay value. The expression is:

[0118]

[0119] Final uncertainty calculation: The uncertainty is obtained by multiplying the square root of the principal diameter element of the posterior covariance by the time delay resolution. The expression is:

[0120]

[0121] in, For posterior covariance The principal diameter element. Used to quantify the reliability of delay estimation, when When this is determined to be a normal scenario with high reliability in latency estimation, When this occurs, it is determined to be a low-reliability scenario; At that time, it was determined to be a medium reliability scenario.

[0122] Step 2.3: Anomaly Identification Generation for Dynamic Thresholds: Based on Time Delay Uncertainty CSI residual energy, Doppler frequency shift f d The system identifies channel anomalies, generates an anomaly identifier and anomaly type, dynamically adjusts the judgment threshold based on the average signal-to-noise ratio of multiple antennas, and forms state features S.

[0123] Multi-antenna average signal-to-noise ratio calculation: Calculate the average signal-to-noise ratio (SNR) of all antennas and subcarriers. SNR reflects CSI quality, and the expression is:

[0124]

[0125] Dynamic threshold Settings: Adjust the anomaly detection threshold based on SNR. When SNR is high, noise is low, so increase the threshold. To reduce the false positive rate, the expression is:

[0126]

[0127] Anomaly detection: CSI residual energy is the difference between the observed and predicted CSI values, and the residual vector is... A is the pilot matrix. If the residual energy exceeds the threshold... If it is not found, it is considered an exception, and an exception flag is output. The expression is:

[0128]

[0129] Anomaly is an anomaly identifier, where 1 = abnormal and 0 = normal.

[0130] The Doppler frequency shift corresponding to the device's movement speed is estimated by the CSI correlation between adjacent frames. The expression is:

[0131]

[0132] in, Let t be the CSI matrix of the t-th frame; The CSI matrix for frame t-1; for The conjugate transpose of ; arccos is the inverse cosine function; It is the Frobenius norm.

[0133] Exception type mapping:

[0134] If Anomaly=1 and It is marked as a new path appearing;

[0135] If Anomaly=1 and Marked as sudden interference;

[0136] If Anomaly=1 and This is marked as device movement.

[0137] The emergence of new paths, sudden interference, and equipment movement are all considered abnormal scenarios, requiring adjustments to pilot configuration to improve latency estimation reliability. Dynamic thresholds can be used to address these issues. Adapt to different SNR scenarios to reduce false detection rate; add device movement type to cover dynamic scenario processing requirements, and output status features after processing. .

[0138] Step 3: Lightweight DQN Intelligent Decision Core: A lightweight deep Q-network (DQN) combining deep separable convolution and fully connected layers is adopted. The network is trained with state features S as input through a multi-objective reward function, and convergence verification is performed to output the optimal pilot configuration action.

[0139] The multi-objective reward function is as follows: the total reward is the weighted sum of the accuracy reward, overhead reward, quality reward and energy consumption reward, and the total weight is 1. The accuracy reward is negatively correlated with the delay uncertainty, the overhead reward is negatively correlated with the pilot density, the quality reward is positively correlated with the signal-to-noise ratio and is halved during anomalies, and the energy consumption reward is negatively correlated with the pilot energy consumption.

[0140] Based on the above, the implementation of step 3 in this embodiment specifically includes the following sub-steps:

[0141] Step 3.1: Lightweight DQN network architecture design: Replace the existing DQN fully connected network with a depthwise separable convolution + fully connected structure to reduce the number of parameters and inference time, define network layer parameters and hyperparameters, and adapt to the computing resources of embedded Wi-Fi devices.

[0142] Network structure:

[0143] Input layer: Fully connected input type, parameter configuration: 4-dimensional state feature S;

[0144] Feature extraction layer: type is depthwise separable convolutional layer, parameter configuration: 3×1 convolutional kernel, 16 output channels, ReLU activation function;

[0145] Hidden layer: Type is fully connected layer, parameter configuration: 32-dimensional neuron, ReLU activation function;

[0146] Output layer: The type is fully connected output. Parameter configuration: 4-dimensional action value, corresponding to 4 resource control actions a1 to a4, linear activation function, and action a includes a1 to a4.

[0147] Hyperparameter optimization: learning rate Reduce the risk of overfitting during small-sample training; increase the capacity of the experience replay pool. Reduce memory usage and adapt to embedded devices; target network update cycle C=50 steps to balance training stability and real-time performance; greedy strategy. The value decreases linearly from 0.8 to 0.1, with initial exploration actions followed by the utilization of optimal actions in the later stages.

[0148] Step 3.2: Design of the energy consumption-aware multi-objective reward function: Set four reward objectives: accuracy, overhead, quality, and energy consumption. The energy consumption objective is the pilot transmission energy consumption. The four objectives are balanced by weight allocation to guide DQN to learn a low-energy-consumption and high-accuracy strategy.

[0149] Multi-objective reward function: The total reward is the weighted sum of the rewards of each sub-objective, expressed as:

[0150]

[0151] Where R represents the total reward; For precision weights; Weighted by cost; For quality weights; The weights are energy consumption weights, and the total weights are 1.

[0152] For accuracy rewards, , The maximum acceptable uncertainty; the lower the uncertainty, the higher the reward.

[0153] As a reward for expenses, , This refers to the pilot density, i.e., the number of pilot symbols per frame. Symbols / frame represent the maximum acceptable pilot overhead; the lower the pilot density, the higher the reward.

[0154] For quality rewards, the higher the SNR and the fewer the anomalies, the higher the reward. 30dB is the ideal SNR. The reward is halved when there are anomalies. The expression is:

[0155]

[0156] As an energy consumption reward, , The pilot power consumption for the current action is obtained from the action-parameter mapping table. The maximum acceptable pilot power consumption is determined by the amount of pilot power consumption; the lower the pilot power consumption, the higher the reward.

[0157] By adding energy consumption incentive items This solves the problem of ignoring energy consumption in existing methods, adapts to battery-powered Wi-Fi terminals, and outputs a total reward value R after processing.

[0158] Step 3.3: Training and Convergence Verification of Lightweight DQN: Using A greedy strategy is used to collect experience, and the DQN is trained by combining experience replay with the target network. The training convergence is verified, and the trained lightweight DQN model is output.

[0159] The training process is as follows:

[0160] 1) Initialize main network parameters Target network parameters And experience replay pool D, initial .

[0161] 2) Experience collection: For each step t:

[0162] Input state S t ,pass Greedy strategy selects action a t S tLet a be the state feature at step t. t The pilot configuration action selected for step t is a discrete action, such as a1 to a4;

[0163] Execute a t Get reward R t and the next state S t+1 R t For step t, perform action a. t The total reward obtained afterward, S t+1 The state features at step t+1;

[0164] Experience (S) t A t ,R t ,S t+1 Store in D.

[0165] 3) Model training: If The experience pool is large enough that 32 experiences (S) are randomly sampled. j A j ,R j ,S j+1 ), calculate the target Q value, the expression is:

[0166]

[0167] Among them, y j The target Q value; R j The total reward for sampling experience; As a discount factor, it balances current and future rewards; The action value output by the target network is used for stable training; j=1,2,…,32 represents 32 experiences randomly sampled from the experience replay pool.

[0168] The loss function is calculated using the following expression:

[0169]

[0170] Where L is the mean squared error loss, which measures the output of the main network. With the target Q value y j Differences; The value of the actions output by the main network;

[0171] Backpropagation: Minimize L using gradient descent to update the main network parameters. ;

[0172] Target network update: every C = 50 steps, let This ensures the stability of training objectives.

[0173] 4) Convergence verification metrics:

[0174] Reward convergence: After 5000 training steps, the total reward R stabilizes at 0.7±0.05, with a fluctuation range of <8%;

[0175] Inference time: Reduced from 10ms in traditional DQN to 4ms, a reduction of 60%;

[0176] Action accuracy: In both normal scenarios and abnormal scenarios such as equipment movement and sudden interference, the accuracy rate of action selection remains above 90%.

[0177] 5) After processing, output the trained lightweight DQN model and the optimal pilot configuration action a.

[0178] Step 4: Energy-Aware Adaptive Resource Control: Execute the optimal pilot configuration action, map the optimal pilot configuration action to pilot configuration parameters, generate and send time-domain orthogonal multi-antenna pilot signals, receive the new CSI matrix fed back from the receiver in the Wi-Fi system, verify the control effect of accuracy improvement rate, overhead reduction rate, and energy consumption reduction rate, and output the new CSI matrix. .

[0179] The mapping of pilot configuration parameters includes pilot density, subcarrier strategy, applicable channel scenario, and pilot power consumption;

[0180] The generation of time-domain orthogonal multi-antenna pilot signals includes: generating pilot signals for each transmit antenna based on the Zadoff-Chu sequence, assigning different dedicated root indices to different transmit antennas, mapping the pilot signals of different transmit antennas to different time-domain segments of the frame, and the receiving end of the Wi-Fi system extracting the corresponding antenna pilot signals through a time-domain window.

[0181] Based on the above, step 4 in this embodiment includes the following sub-steps:

[0182] Step 4.1: Multi-antenna action-parameter mapping: Map the discrete actions a1 to a4 output by DQN to specific pilot configuration parameters to adapt to different channel scenarios, including normal scenario, low reliability scenario, medium reliability scenario, main path defined scenario and abnormal scenario.

[0183] The action-parameter mapping is as follows:

[0184] action Pilot density Subcarrier strategy: uniformly select 10 subcarriers; multi-antenna pilot allocation: all antennas share pilots; applicable channel scenario: normal scenario. Pilot power consumption ;

[0185] action Pilot density Subcarrier strategy: Select all 52 subcarriers; Multi-antenna pilot allocation: Independent pilot for each antenna; Applicable channel scenario: Low reliability scenario. Pilot power consumption ;

[0186] action Pilot density Subcarrier strategy: main path subcarrier ±2 frequency bands; multi-antenna pilot allocation: multi-antenna joint pilot; applicable channel scenarios: clear main path. Stable, pilot power consumption ;

[0187] action Pilot density Subcarrier strategy: Eliminate subcarriers with SNR < 10dB; Multi-antenna pilot allocation: Reduce pilot signal usage by eliminating interfering antennas; Applicable channel scenarios: Sudden interference / equipment movement, pilot power consumption .

[0188] The pilot configuration strategy for medium reliability scenarios is as follows: pilot density is between and This approach, which can ensure accurate delay estimation by focusing on the main path subcarrier, avoids excessive overhead or energy waste due to excessive pilot density, and is the preferred choice for medium reliability scenarios. It should be noted that fine-tuning can be done by combining SNR and anomaly indicators. For example, if SNR > 20dB and Anomaly = 0, a downgrade can be selected. To further reduce costs; if SNR < 10dB, an upgrade option can be selected. This enhances noise immunity, thereby enabling adaptive flexibility.

[0189] The processed multi-antenna pilot configuration parameters are output as inputs for the generation and transmission of multi-antenna pilot signals, generating physical layer pilot signals.

[0190] Step 4.2: Multi-antenna pilot signal generation and transmission: Based on the Zadoff-Chu sequence, generate time-domain orthogonal multi-antenna pilot signals to avoid inter-antenna interference, ensure accurate demodulation at the receiver, output pilot signals and transmit them, and the receiver feeds back a new CSI matrix. .

[0191] Multi-antenna Zadoff-Chu sequence generation, the expression is:

[0192]

[0193] in, Let N be the pilot sequence for the x-th transmitting antenna; k is the sequence index; N is the number of antennas. ZC=13 is the Zadoff-Chu sequence length, adapted to the 802.11ax pilot format; For antenna-specific root indices, tx1: u=1, tx2: u=3, ensure that the correlation between different antenna sequences is <0.1 to avoid interference.

[0194] Multi-antenna pilot time-domain orthogonal mapping: The Zadoff-Chu sequences of different antennas are mapped to different time-domain segments of the frame: pilot tx1 is mapped to the first half of the frame, and pilot tx2 is mapped to the second half of the frame; the receiver extracts the pilot signals of the corresponding antennas through the time-domain window to eliminate inter-antenna interference.

[0195] By assigning dedicated root indexes to different transmitting antennas It employs time-domain orthogonal mapping to avoid interference between multiple antenna pilots, outputting time-domain orthogonal multi-antenna pilot signals, which are transmitted to the terminal via the Wi-Fi RF module. After receiving and demodulating the signals, the terminal generates a new CSI matrix. Feedback is sent to the sending end for control effect verification.

[0196] Step 4.3: Multi-dimensional control effect verification: Calculate the control effect from three dimensions: accuracy improvement, overhead reduction and energy consumption reduction. Verify whether the current action has achieved the accuracy-overhead-energy consumption balance target. If it has, enter the closed-loop feedback. If it has not, re-trigger the DQN decision and select a new action.

[0197] Improvement rate of computational accuracy The expression is:

[0198]

[0199] in, To control the uncertainty before it is controlled; To mitigate post-control uncertainties, feedback is provided via the terminal. Calculated.

[0200] Computational cost reduction rate The expression is:

[0201]

[0202] in, Pilot density before control; This refers to the pilot density after control.

[0203] Calculate the energy consumption reduction rate The expression is:

[0204]

[0205] in, Pilot power consumption before control; The pilot power consumption after control is obtained from the action-parameter mapping table.

[0206] Performance Standard: Precision Improvement Rate ; Cost reduction rate Energy consumption reduction rate If all three criteria are met, the control is deemed effective.

[0207] The processed output control performance indicators and the new CSI matrix from the terminal feedback are also included. Control effectiveness indicators include , and .

[0208] Step 5: Closed-loop optimization feedback of dynamic threshold: for the new CSI matrix Repeat step 1 to eliminate errors and obtain a clean new CSI matrix after preprocessing. Update the noise variance and prior covariance of the multi-antenna Bayesian model and the network parameters of the lightweight DQN to perform closed-loop optimization of dynamic channel adaptation.

[0209] The update of the multi-antenna Bayesian model includes: re-estimating the noise variance of the model based on clean new CSI matrix samples, with the number of samples increasing with each feedback; if the amplitude deviation of the multi-antenna main diameter exceeds 10%, the prior covariance of the model is increased, otherwise the prior covariance remains unchanged.

[0210] The update of network parameters for lightweight DQN includes: calculating new state features based on a clean new CSI matrix, generating new experience by combining pilot configuration actions, total reward and next state, and storing it in the experience replay pool; randomly sampling experience from the replay pool, repeatedly training the network, updating the parameters of the main network and the target network, and verifying the action selection accuracy in the device movement scenario to confirm the update effect.

[0211] Based on the above, step 5 in this embodiment specifically includes the following sub-steps:

[0212] Step 5.1: Multi-antenna feedback data preprocessing: Preprocessing the new CSI matrix fed back from the terminal. Repeat the preprocessing steps from step 1. Perform multi-antenna joint SFO calibration in step 1.2 to eliminate SFO drift, and perform multi-antenna phase noise KF smoothing in step 1.3 to eliminate phase error, resulting in a clean new CSI matrix. .

[0213] Step 5.2: Multi-antenna Bayesian model parameter update: update the parameters of the new samples. Incorporate Bayesian model to update noise variance and prior covariance This improves the model's adaptability to dynamic channels.

[0214] Multi-antenna noise variance renew:

[0215] Based on the accumulated M clean samples, the noise variance is re-estimated. The more samples, the more accurate the estimation. The expression is:

[0216]

[0217] Where M is the cumulative number of samples, and M increases by 1 for each feedback. This is the m-th clean CSI sample; Let be the posterior mean of the m-th sample.

[0218] Multi-antenna prior covariance update: The prior covariance is dynamically adjusted based on principal diameter consistency. If the principal diameter deviation of the multi-antenna is >10%, or if spatial consistency is disrupted due to equipment movement, the prior covariance is increased. To enhance model robustness and allow for larger multipath amplitude fluctuations, the expression is:

[0219]

[0220] in, For the updated prior covariance; For the prior covariance before updating; Main diameter amplitude;

[0221] By dynamically adjusting the prior covariance based on the consistency of the main diameter of multiple antennas, overfitting of the model in dynamic scenarios is avoided, and the optimized multi-antenna Bayesian model is output after processing. More accurate More robust; latency uncertainty reduce.

[0222] Step 5.3: Lightweight DQN Model Update: Store new experiences in the experience replay pool, retrain DQN periodically, update network parameters, and improve the model's accuracy in selecting actions in dynamic scenes.

[0223] Update process:

[0224] New experience generation: based on Calculate the new state By combining the actions from step 4, the total reward, and the next state, new experience is generated. Add to experience replay pool D;

[0225] Periodic retraining: Every 50 frames, randomly sample 1000 experiences from the experience replay pool D, repeat the training process in step 3.3, and update the main network parameters. and target network parameters ;

[0226] Convergence verification: In mobile device scenarios, the action selection accuracy improved from 80% before the update to over 88%.

[0227] The updated lightweight DQN model is then output after processing, serving as the decision model for the next step (Step 3) to improve the accuracy of action selection in dynamic scenarios.

[0228] Example 2:

[0229] like Figure 2 As shown, this embodiment provides a Wi-Fi channel state information delay estimation system, including:

[0230] The multi-antenna data preprocessing module is used to collect raw channel state information data from multiple antennas, eliminate hardware errors through joint calibration and filtering of multiple antennas, and output the preprocessed multi-antenna CSI matrix.

[0231] The state feature extraction module is used to receive the preprocessed multi-antenna CSI matrix, calculate the delay value, delay uncertainty and channel anomaly identifier through Bayesian estimation of multi-antenna fusion, and form state features;

[0232] The intelligent decision-making module receives state features, performs decision analysis through a lightweight deep Q-network, and outputs the optimal pilot configuration action.

[0233] The resource control and verification module is used to receive the optimal pilot configuration action, generate and send multi-antenna pilot signals, receive the new CSI matrix fed back by the terminal, and verify the control effect.

[0234] The closed-loop optimization module is used to receive the new CSI matrix and update the Bayesian model parameters in the state feature extraction module and the network parameters in the intelligent decision-making module.

[0235] The multi-antenna data preprocessing module includes a crystal oscillator offset calibration unit and a Kalman filter smoothing unit. The crystal oscillator offset calibration unit calculates the crystal oscillator offset frequency by fusing the phase difference of all antenna pairs and eliminates linear phase drift. The Kalman filter smoothing unit performs iterative prediction and correction by fusing multi-antenna observations to eliminate random phase noise.

[0236] The state feature extraction module includes a Bayesian delay estimation unit and a dynamic anomaly identification unit. The Bayesian delay estimation unit constructs a joint Bayesian model containing principal path consistency constraints and calculates the fused delay value. The dynamic anomaly identification unit adaptively adjusts the judgment threshold based on the signal-to-noise ratio and generates anomaly types.

[0237] The intelligent decision-making module includes a lightweight DQN structure and a multi-objective reward function unit. The lightweight DQN structure uses a combination of depthwise separable convolution and fully connected layers. The multi-objective reward function unit includes a weighted combination of accuracy reward, overhead reward, quality reward and energy reward.

[0238] The resource control and verification module includes an action-parameter mapping unit and a time-domain orthogonal pilot generation unit. The action-parameter mapping unit converts discrete actions into pilot density and subcarrier policy parameters, while the time-domain orthogonal pilot generation unit generates orthogonal pilot signals for different antennas based on the Zadoff-Chu sequence.

[0239] The closed-loop optimization module includes a Bayesian model update unit and a lightweight DQN update unit. The Bayesian model update unit updates the noise variance and prior covariance based on accumulated samples, while the lightweight DQN update unit periodically retrains the network parameters through empirical replay.

Claims

1. A method for estimating delay of Wi-Fi channel state information, characterized in that, The method comprises the following steps: S1: preprocessing the collected multi-antenna CSI data to eliminate the hardware errors of crystal oscillator offset, radio frequency phase noise and multi-antenna phase inconsistency, and obtaining a multi-antenna CSI matrix; S2: based on the multi-antenna CSI matrix, constructing a multi-antenna joint Bayesian model, extracting each antenna main path and performing consistency detection, calculating the fusion time delay value, time delay uncertainty and channel anomaly identifier, and forming a state feature; S3: using a lightweight DQN combining deep separable convolution and full connection, taking the state feature as input, training the DQN through a multi-objective reward function, and performing convergence verification, and outputting an optimal pilot configuration action; S4: executing the optimal pilot configuration action and mapping it to pilot configuration parameters, generating and sending multi-antenna pilot signals with time domain orthogonality, receiving feedback new CSI matrix, and outputting the new CSI matrix after verification; S5: repeating the preprocessing of step S1 on the new CSI matrix to obtain the preprocessed new CSI matrix, updating the network parameters of the multi-antenna Bayesian model and the lightweight DQN, and performing closed-loop optimization on the dynamic channel adaptation.

2. The method of claim 1, wherein, The elimination of the crystal oscillator offset in step S1 comprises: calculating the average phase difference of a single antenna pair between adjacent frames, fusing the phase differences of all antenna pairs to determine the crystal oscillator offset frequency, and inversely eliminating the linear phase drift of the same subcarrier of adjacent frames based on the crystal oscillator offset frequency to obtain the multi-antenna true phase after eliminating the crystal oscillator offset; The elimination of the radio frequency phase noise comprises: smoothing through multi-antenna phase noise Kalman filtering, taking the weighted average of the multi-antenna phase as the observation value, and the weight of the weighted average being the inverse of the variance of each antenna phase, iteratively predicting and correcting to eliminate the random phase noise introduced by the radio frequency link; The elimination of the hardware error of multi-antenna phase inconsistency comprises: during the elimination of the crystal oscillator offset and the Kalman filtering smoothing, the phase inconsistency among the multi-antenna is eliminated together by fusing the multi-antenna observation data.

3. The method of claim 1, wherein, The construction of the multi-antenna joint Bayesian model in step S2 comprises: constructing a multi-antenna joint Bayesian model based on the linear mapping relationship between the frequency domain CSI and the multipath time delay, and the multi-antenna joint Bayesian model contains multi-antenna prior constraints, the prior constraints including the multi-antenna main path amplitude deviation and the sparse prior of the multipath time delay amplitude vector, and the sparse prior is used to ensure that only the main path has a non-zero amplitude.

4. The method of claim 1, wherein, The calculation of the channel anomaly identifier in step S2 comprises: A1: calculating the average signal-to-noise ratio (SNR) of all antennas and subcarriers, and setting a dynamic threshold according to the SNR; A2: calculating the residual energy of the CSI observation value and the predicted value, and marking the anomaly if the residual energy exceeds the dynamic threshold; A3: estimating the Doppler shift through the correlation of adjacent frame CSIs, and determining the anomaly type in combination with the residual energy, the time delay uncertainty and the Doppler shift.

5. The method of claim 1, wherein, The structure design of the lightweight DQN in step S3 comprises: replacing the full connection network of the existing DQN with a deep separable convolution+full connection structure, and defining the network layer parameters and hyperparameters; The network structure comprises: Input layer: full connection input type; Feature extraction layer: deep separable convolution layer type; Hidden layer: full connection layer type; Output layer: full connection output type.

6. The method of claim 1, wherein, The pilot configuration parameter mapping in the step S4 includes pilot density, subcarrier strategy, applicable channel scenario and pilot energy consumption; The generation of the time-domain orthogonal multi-antenna pilot signal includes: generating a pilot signal of each transmitting antenna based on a Zadoff-Chu sequence, assigning different transmitting antennas with different exclusive root indexes, and mapping the pilot signals of different transmitting antennas to different time-domain segments of a frame, and a receiving end of the Wi-Fi system extracts the pilot signal of the corresponding antenna through a time-domain window.

7. The method of claim 1, wherein, The multi-antenna Bayesian model updating in the step S5 includes: re-estimating the noise variance of the model based on the preprocessed new CSI matrix sample, and increasing the sample number with each feedback; if the multi-antenna main path amplitude deviation exceeds a set value, the prior covariance of the model is expanded, otherwise the prior covariance remains unchanged; The network parameter updating of the lightweight DQN includes: calculating a new state feature based on the preprocessed new CSI matrix, combining the pilot configuration action, the total reward and the next state to generate a new experience, and storing the experience in an experience replay pool; randomly sampling the experience from the experience replay pool, repeatedly training the network, updating the main network and target network parameters, and verifying the action selection accuracy in the device moving scenario.

8. A system for Wi-Fi channel state information latency estimation, the system comprising: The system for implementing the method of claim 1 comprises: A multi-antenna data preprocessing module is configured to collect multi-antenna original channel state information data, eliminate hardware errors through multi-antenna joint calibration and filtering, and output preprocessed multi-antenna CSI matrix; A state feature extraction module is configured to receive the preprocessed multi-antenna CSI matrix, calculate the time delay value, time delay uncertainty and channel anomaly identifier through multi-antenna fusion Bayesian estimation, and form the state feature; An intelligent decision-making module is configured to receive the state feature, perform decision analysis through the lightweight deep Q network, and output the optimal pilot configuration action; A resource control and verification module is configured to receive the optimal pilot configuration action, generate and send the multi-antenna pilot signal, receive the new CSI matrix fed back by the terminal and verify the control effect; A closed-loop optimization module is configured to receive the new CSI matrix, update the Bayesian model parameters in the state feature extraction module and the network parameters in the intelligent decision-making module.

9. The Wi-Fi channel state information time delay estimation system according to claim 8, wherein The multi-antenna data preprocessing module includes a crystal oscillator offset calibration unit and a Kalman filter smoothing unit, the crystal oscillator offset calibration unit calculates the crystal oscillator offset frequency and eliminates the phase linear drift by fusing the phase difference of all antenna pairs, and the Kalman filter smoothing unit iteratively predicts and corrects to eliminate random phase noise by fusing the multi-antenna observation values; The state feature extraction module includes a Bayesian time delay estimation unit and a dynamic anomaly identification unit, the Bayesian time delay estimation unit constructs a joint Bayesian model containing main path consistency constraints and calculates the fused time delay value, and the dynamic anomaly identification unit adaptively adjusts the judgment threshold based on the signal-to-noise ratio and generates the anomaly type. The intelligent decision module comprises a lightweight DQN structure and a multi-objective reward function unit, the lightweight DQN structure adopts a combination of a deep separable convolution and a full connection layer, and the multi-objective reward function unit comprises a weighted combination of an accuracy reward, an overhead reward, a quality reward and an energy consumption reward; The resource control and verification module comprises an action-parameter mapping unit and a time domain orthogonal pilot generation unit, the action-parameter mapping unit converts discrete actions into pilot density and subcarrier strategy parameters, and the time domain orthogonal pilot generation unit generates orthogonal pilot signals of different antennas based on a Zadoff-Chu sequence; The closed loop optimization module comprises a Bayesian model updating unit and a lightweight DQN updating unit, the Bayesian model updating unit updates noise variance and prior covariance based on cumulative samples, and the lightweight DQN updating unit regularly re-trains network parameters through experience replay.

Images