A joint spectrum recovery and target positioning method based on wi-fi channel state information

Abstract

This invention discloses a joint spectrum recovery and target localization method based on Wi-Fi channel state information, belonging to the field of wireless communication sensing and positioning technology. Addressing the problems of low joint spectrum resolution and large positioning errors in existing Wi-Fi positioning methods, this method preprocesses the channel state information from multiple antennas and frequencies, standardizing it using training set statistics. Then, it inputs this information into a joint spectrum recovery network composed of a convolutional encoder-decoder, a two-dimensional bidirectional state space module, and a delay feature enhancement module to recover the angle-of-arrival-time-of-flight joint spectrum. During training, a logarithmic transformation is introduced, employing a model constrained by logarithmic spatial error, original spatial error, and peak error. During inference, the network output is restored to the original spectral space, and the angle of arrival and time of flight are obtained through peak search and path filtering to estimate the target position. This invention improves the quality of joint spectrum recovery, reduces spectral peak shift and abnormal positioning errors, and enhances the stability of Wi-Fi positioning.

Description

Technical Field

[0001] This invention relates to the field of wireless communication sensing and positioning technology, specifically to a method for joint angle-of-arrival-time-of-flight spectrum recovery and target position estimation based on Wi-Fi channel state information, which is particularly suitable for target positioning tasks under conditions of limited bandwidth and sparse channel sampling. Background Technology

[0002] Wi-Fi signals can be used for indoor people positioning, security monitoring, smart homes, and robot navigation. Traditional methods mostly determine target location based on received signal strength, but signal strength is easily affected by environmental changes and provides limited information. Channel state information, in addition to recording signal amplitude, also includes phase changes at different receiving antennas and frequency sampling points, thus reflecting the different paths the wireless signal takes during propagation. Existing positioning methods typically estimate the signal's direction of arrival and propagation time based on this amplitude and phase information to further determine the target location.

[0003] The angle of arrival (AHA) indicates the direction from which the signal arrives at the receiver, while the time of flight (TOF) represents the time it takes for the signal to travel from the transmitter to the receiver. Placing the AHA and TOF on the same two-dimensional coordinate system yields the joint AHA-TOF spectrum. Peaks in the joint spectrum typically correspond to propagation paths with higher energy, and the location of these peaks reflects the angle and time delay of the relevant path. In practical applications, ordinary Wi-Fi signals have narrow bandwidths and a limited number of collectable frequency points. Furthermore, indoor environments present challenges such as reflection, multipath propagation, and hardware phase deviations. Therefore, the recovered joint spectrum is prone to issues such as unclear peaks, positional shifts, or unstable results.

[0004] Currently, existing methods employ convolutional encoder-decoder architectures to recover the angle-of-arrival-time-of-flight joint spectrum based on limited channel state information. While these networks are relatively convenient for extracting local features, they do not adequately consider the connections between geographically distant regions in the spectrum, potentially leading to peak shifts, spurious peaks, or incomplete local structure recovery in complex multipath environments. Furthermore, the numerical range of the joint spectrum is typically large; if trained directly in the original spectral space, the model is easily affected by high-energy regions, making it difficult to simultaneously capture details in weak response regions and near peaks.

[0005] Therefore, existing methods still need to further improve their ability to model local time delay features and the global correlation between angle of arrival and time of flight, and simultaneously constrain the overall shape of the spectrum, the original spectral values, and the peak positions during training to meet the requirements of subsequent position calculation. Summary of the Invention

[0006] The purpose of this invention is to provide a joint spectrum recovery and target localization method based on Wi-Fi channel state information, so as to solve the problems in the prior art that are difficult to stably recover the angle-of-arrival-time-of-flight joint spectrum due to limitations in signal bandwidth and frequency sampling conditions, as well as the problems of large spectral peak shift, false peaks and localization errors.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A joint spectrum recovery and target localization method based on Wi-Fi channel state information includes the following steps:

[0009] Step 1: Obtain channel state information at different receiving antennas and different frequency sampling points at the receiving end, organize and preprocess the channel state information to obtain network input features;

[0010] Step 2: Standardize the network input features based on the statistical parameters obtained from the input features of the training set;

[0011] Step 3: The standardized network input features are fed into the joint spectrum recovery network, which is based on a convolutional encoder-decoder and includes a two-dimensional bidirectional state space module and a time delay feature enhancement module. The two-dimensional bidirectional state space module is used to perform sequence modeling along the row and column directions of the two-dimensional features, and the time delay feature enhancement module is used to extract local change features related to propagation delay. The joint spectrum recovery network then obtains the angle-of-arrival-time-of-flight joint spectrum.

[0012] Step 4: Perform peak search on the angle-of-arrival-time-of-flight joint spectrum, determine the target-related propagation path from the candidate spectrum peaks, and obtain the angle of arrival and time of flight corresponding to the target-related propagation path;

[0013] Step 5: Based on the angle of arrival, flight time, known distance between the transmitter and receiver, and pre-established spatial geometric relationships, obtain the position of the target to be located.

[0014] A joint spectrum recovery and target localization device based on Wi-Fi channel state information, comprising:

[0015] The data processing module is used to acquire channel state information of the receiver at different receiving antennas and different frequency sampling points, organize and preprocess the channel state information, and use the statistical parameters obtained in the training phase to complete the standardization process to obtain network input features.

[0016] The joint spectrum recovery module is used to feed the standardized network input features into the joint spectrum recovery network to obtain the angle-of-arrival-time-of-flight joint spectrum. The joint spectrum recovery network includes a convolutional encoder-decoder, a two-dimensional bidirectional state space module, and a time delay feature enhancement module. The two-dimensional bidirectional state space module is used to establish long-distance correlations along the row and column directions of the two-dimensional features, and the time delay feature enhancement module is used to extract local change features related to propagation time delay.

[0017] The parameter extraction module is used to perform peak search on the angle-of-arrival-time-of-flight joint spectrum, sort the candidate spectral peaks according to their spectral values, and determine the target-related propagation path according to preset path filtering conditions, thereby obtaining the angle of arrival and time of flight corresponding to the target-related propagation path.

[0018] The position calculation module is used to calculate the two-dimensional position coordinates of the target to be located according to the angle of arrival, flight time and the known distance between the transmitter and receiver, and in accordance with the pre-established spatial geometric relationship.

[0019] The model training module is used to perform a logarithmic transformation on the angle-of-arrival-time-of-flight joint spectral labels, and construct a joint loss function based on the logarithmic spatial error, the original spatial error, the peak region error, and the peak coordinate error. The parameters of the joint spectral reconstruction network are iteratively updated using the joint loss function to obtain the trained joint spectral reconstruction network.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] (1) The present invention adds a two-dimensional bidirectional state space module to the convolutional encoder-decoder to process two-dimensional features along the row and column directions respectively, thereby supplementing the problem of insufficient modeling of long-distance feature association by the convolutional structure and improving the recovery effect of the overall structure of the joint spectrum.

[0022] (2) The present invention sets up a time delay feature enhancement module to extract local changes in the frequency sampling direction and fuse the time delay response in the frequency domain to enhance the time delay features near the target spectral peak.

[0023] (3) Before training, the input features are standardized using the mean and standard deviation of the training set, and the joint spectral labels are logarithmically transformed to reduce the impact of input feature differences and spectral dynamic range on training. During training, the logarithmic spatial error, the original spatial error, and the peak correlation error are calculated simultaneously to ensure that the recovery result takes into account the overall spectral shape, spectral value magnitude, and peak position.

[0024] (4) This invention determines the target-related propagation path for location calculation by performing multi-peak search and path screening on the recovered angle-of-arrival-time-of-flight joint spectrum, and then completes the two-dimensional coordinate calculation by combining the angle of arrival, time of flight and known distance between the transmitter and receiver. This can reduce the impact of false spectral peaks and abnormal paths on the positioning results and improve the stability of target position estimation. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall system of the joint spectrum recovery and target localization method based on Wi-Fi channel state information according to the present invention.

[0026] Figure 2 This is a diagram of the overall structure of the joint spectrum recovery network in this invention.

[0027] Figure 3 This is a structural diagram of the delay feature enhancement module in this invention.

[0028] Figure 4 This is a structural diagram of the two-dimensional bidirectional state space module in this invention.

[0029] Figure 5 This is a flowchart of peak selection, target location calculation, and error evaluation in this invention. Detailed Implementation

[0030] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0031] The overall process of this embodiment is as follows: Figure 1 As shown, the process includes, in sequence, channel state information acquisition, input feature processing and standardization, angle-of-arrival-time-of-flight joint spectrum recovery, target-related path filtering, and target location calculation.

[0032] The positioning system in this embodiment includes a Wi-Fi signal transmitter, a Wi-Fi signal receiver, and a target to be located. The transmitter transmits detection signals at multiple frequency sampling points, and the receiver receives direct signals arriving directly from the transmitter, as well as reflected signals arriving at the receiver after being reflected by the target. The receiver obtains channel state information at different receiving antennas and different frequency sampling points based on the received signals.

[0033] In this embodiment, the receiver uses multiple receiving antennas to acquire channel state information. The channel state information includes the complex channel response corresponding to each frequency sampling point and each receiving antenna. The complex channel response includes amplitude and phase. The phase change between frequency sampling points reflects the propagation delay, and the phase difference between different receiving antennas reflects the signal arrival direction. Together, they constitute the information required for joint spectrum recovery.

[0034] After acquiring the channel state information, the data is arranged according to the receiving antenna dimension and frequency sampling dimension, and the real part, imaginary part, or related features obtained from their transformation are organized into network input features. In one embodiment, the size of the network input features is B×4×121×281, where B represents the batch sample size, 4 represents the number of input feature channels, 121 corresponds to the time-of-flight sampling dimension, and 281 corresponds to the angle-of-arrival sampling dimension. In this embodiment, the joint spectrum recovery network has 4 input channels and 1 output channel.

[0035] The numerical ranges of different samples and different feature channels may vary significantly. To reduce the impact of this variation on network training, the mean and standard deviation are first calculated based on the input features of the training set, and then the input features are standardized according to the following formula:

[0036]

[0037] Where x represents the input features before standardization. This represents the standardized input features. and These represent the mean and standard deviation obtained from the training set, respectively. A small constant is set to prevent the denominator from being zero. The mean and standard deviation obtained in the training phase will continue to be used in the testing and actual positioning phases, and will not be recalculated based on the test data.

[0038] like Figure 2 As shown, the standardized input features are fed into the joint spectrum recovery network. The joint spectrum recovery network adopts a convolutional encoder-decoder structure, including an input mapping module, a three-level coding unit, a time delay feature enhancement module, a bottleneck feature processing module, a three-level decoding unit, and a joint spectrum output module.

[0039] The input mapping module is used to perform channel mapping and shallow feature extraction. In this embodiment, this module includes 3×3 convolution, group normalization, and SiLU activation function, where group normalization can reduce the impact of batch size variations on feature statistics.

[0040] After the input feature map is completed, it passes through three levels of encoding units. Each level of encoding unit includes a residual convolutional block and downsampling. The residual convolutional block is used to extract local features, and downsampling can reduce the resolution of the feature map and expand the receptive field. The further the encoding level, the smaller the feature map size and the larger the number of channels.

[0041] The residual convolutional block consists of a main branch and a shortcut branch. The main branch performs convolution, normalization, and activation operations sequentially, while the shortcut branch retains the input information. When the number of input channels and the number of output channels are not equal, a 1×1 convolution is used in the shortcut branch to perform channel matching. The output of the residual convolutional block can be represented as:

[0042]

[0043] in, Indicates input features, This represents the feature transformation of the main branch of the convolution. This represents the mapping of shortcut branches.

[0044] A time-delay feature enhancement module is set on the encoded deep features. The structure of the time-delay feature enhancement module is as follows: Figure 3 As shown, this module includes a first convolutional branch, a second convolutional branch, a latency-aware branch, and a gated fusion unit.

[0045] The first convolutional branch performs convolution along the direction corresponding to the frequency sampling in the input features, used to extract local features that vary with frequency. The second convolutional branch performs convolution along a direction orthogonal to this direction, used to preserve the local structure in another direction. The two convolutional branches extract local changes in two directions respectively, so that the spectral peaks and their neighborhood structures are preserved more completely.

[0046] The delay-aware branch performs a Fourier transform along the feature dimension corresponding to the frequency sampling, transforming the features from the original representation domain to the frequency domain, and obtains the features related to the propagation delay based on the transformed response. The output of the delay-aware branch is restored to the same size as the input features and fused with the outputs of the two convolutional branches, thereby highlighting the frequency variations and local spectral peaks related to the propagation delay.

[0047] The outputs of the first convolutional branch, the second convolutional branch, and the delay-aware branch are processed by a gated fusion unit. Let the outputs of the three branches be... , and The gating weight can then be expressed as:

[0048]

[0049] in, Indicates feature splicing, Represents a convolution mapping. This represents the Sigmoid function. The fusion result is represented as:

[0050]

[0051] in This represents the characteristics after the branches are merged. This indicates element-wise multiplication. This represents the input to the time-delay feature enhancement module. The gating weights determine the fusion ratio between the enhanced features and the original features, and are adjusted according to changes in the features of the input samples.

[0052] After the time-delay feature enhancement is completed, the features enter the bottleneck layer. The bottleneck layer consists of residual convolutional blocks, a two-dimensional bidirectional state space module, and a time-delay feature enhancement module. The residual convolutional blocks extract local features, and the two-dimensional bidirectional state space module establishes connections between distant locations along two directions, supplementing the shortcomings of ordinary convolution in modeling global information.

[0053] The structure of a two-dimensional bidirectional state-space module is as follows: Figure 4 As shown. For input features Feature sequences are constructed along the row and column directions, respectively. When processing along the row direction, each row is regarded as a sequence of length W; when processing along the column direction, each column is regarded as a sequence of length H.

[0054] Row direction features and column direction features can be represented as follows:

[0055]

[0056]

[0057] in, and These represent the state-space modeling processes along the row and column directions, respectively. After processing, the outputs in both directions are restored to a two-dimensional feature form, and then fused through concatenation and convolution mapping.

[0058]

[0059] in, This represents a convolutional mapping. After modeling the state space in both row and column directions, the module establishes long-distance associations in the angular and temporal dimensions while preserving the two-dimensional arrangement relationship.

[0060] The bottleneck layer outputs into the decoder. The decoder contains three levels of decoding units, which progressively recover the spatial dimensions of the feature maps. Each level first performs upsampling through bilinear interpolation, then concatenates it with the shallow features of the corresponding coding layer, and finally completes the fusion using convolutional and residual convolutional blocks.

[0061] A skip connection is established between the encoder and decoder to directly pass shallow features retained during the encoding stage to the corresponding decoding unit. This compensates for the spectral peak details and local edge information lost during continuous downsampling, enabling the decoding stage to utilize both deep semantic features and shallow structural features simultaneously.

[0062] The decoded features are channel-mapped by the joint spectrum output module, and non-negativity constraints are applied using the Softplus function. The Softplus function is expressed as:

[0063]

[0064] This function ensures that the output spectral value is not less than zero. The network's final output is a joint angle-of-arrival-time-of-flight spectrum with dimensions of B×1×121×281.

[0065] Because the original joint spectrum has a large numerical range, if training is performed directly in the original spectral space, high-amplitude regions tend to occupy a large proportion of the loss function. Therefore, a logarithmic transformation is performed on the labels of the original joint spectrum before training:

[0066]

[0067] Where y represents the original joint spectrum label, This represents the logarithmic space label. The network is trained in the logarithmic space, and during inference, the inverse exponential transform is used to recover the original joint spectrum.

[0068]

[0069] in, This represents the logarithmic space output of the network. This represents the restored joint spectrum.

[0070] Before the inverse exponential transform, the network output is cropped according to the range of values ​​of the training set labels, and non-numerical and infinite results are processed to reduce the impact of outlier spectral values ​​on subsequent peak search.

[0071] This embodiment uses a joint loss function to train the joint spectral recovery network. The joint loss function includes log-space mean squared error, original space mean squared error, original space Charbonnier loss, peak region weighted error, and peak coordinate error, which can be expressed as:

[0072]

[0073] in, to These represent the weights corresponding to each loss term.

[0074] Log-space mean square error is used to constrain the overall distribution of the joint spectrum:

[0075]

[0076] The original spatial mean square error is used to directly reduce the pixel-wise error of the reconstructed joint spectrum:

[0077]

[0078] The original space Charbonnier loss is expressed as:

[0079]

[0080] in, It is a relatively small constant. This loss is less sensitive to anomalous pixels than the ordinary mean square error, which can reduce the impact of individual anomalous spectral values ​​on model training.

[0081] The peak region weighting error generates a weight map based on the true joint spectrum, giving higher weights to pixels near the spectral peak. In this embodiment, the true joint spectrum is normalized by dividing it by its maximum spectral value, and a continuous weight map is generated based on the normalization result. The peak coordinate error is estimated by using a differentiable method to determine the positions of the predicted and true spectral peaks, and the deviation between them is calculated.

[0082] During training, the weights of each loss term are adjusted in stages. At the beginning of training, logarithmic spatial loss is used to allow the network to learn the overall shape of the joint spectrum; then the weights of the original spatial loss and peak correlation loss are increased to further correct the magnitude of the spectral values ​​and the peak positions.

[0083] In one embodiment, the network is trained using the AdamW optimizer, with the learning rate varying according to a cosine annealing strategy. During testing, model parameters are calculated using an exponential moving average to reduce the impact of parameter fluctuations in the later stages of training. During training, the original spatial root mean square error, mean absolute error, and peak position error are recorded, and the model with the smaller original spatial root mean square error is selected for joint spectral prediction.

[0084] In one specific embodiment, the training set contains 7050 samples, the test set contains 1410 samples, the network input size is 4×121×281, and the output size is 1×121×281. The training process can be completed in a Python or PyTorch environment. After training, the prediction results of the test set are saved as a MAT format file for subsequent peak search and target location calculation.

[0085] After the joint spectrum recovery network outputs the two-dimensional angle-of-arrival-time-of-flight joint spectrum, according to... Figure 5 The illustrated process includes candidate peak extraction, path filtering, location calculation, and error analysis. Peak search utilizes a pre-defined angle of arrival (Angle of Arrival) and time-of-flight grid. In one embodiment, the Angle of Arrival search range is 20° to 160°, with intervals of 0.5°; the Time-of-Flight search range is... to , interval is s.

[0086] A peak search was performed on the predicted joint spectrum to extract 10 candidate spectral peaks, and the angle of arrival, time of flight, and spectral value of each candidate peak were recorded. The candidate peaks were then arranged in descending order of spectral value.

[0087] After sorting, the top 5 candidate peaks are sent to the path selection process. In this embodiment, candidate peaks with an angle of arrival of 90° or within a preset tolerance range near 90° and a flight time close to the preset direct path flight time are identified as direct path peaks. For the remaining candidate peaks, the first one with an angle of arrival between 20° and 160°, whose angle of arrival does not meet the direct path angle condition, and whose flight time differs from the direct path flight time by more than [amount missing]. The candidate peak of s serves as the target-related propagation path.

[0088] The path selection results include, in order, the angle of arrival and flight time of the direct path, and the angle of arrival and flight time of the target-related propagation path. When calculating the target position, the angle of arrival and flight time corresponding to the target-related propagation path are selected as the position calculation parameters.

[0089] Assume the transmitter and receiver are located on the same coordinate axis, and the known distance between them is... The angle of arrival of the target-related reflection path is Flight time is The electromagnetic wave propagation speed is c. First, calculate the total length of the target-related propagation path:

[0090]

[0091] Then, based on the geometric relationship between the transmitter, the target, and the receiver, the distance from the target to the receiver is calculated. :

[0092]

[0093] And according to and the angle of arrival The target coordinates (x, y) are obtained:

[0094]

[0095]

[0096] In this embodiment, the known distance between the transmitter and receiver The value is set to 7 m, and the electromagnetic wave propagation speed c is set to 299,792,458 m / s; in actual use, It should be set according to the actual installation distance between the transmitter and receiver.

[0097] To verify the positioning results, multiple estimated coordinates corresponding to the same actual location are grouped together. In this embodiment, every 30 estimated coordinates correspond to one actual location. Let the estimated location be... The corresponding real location is The positioning error is then expressed as:

[0098]

[0099] All valid positioning errors were statistically analyzed to obtain the mean, median, standard deviation, 75th percentile, 90th percentile, 95th percentile, 99th percentile, and maximum error. The number of samples with errors greater than 0.2 m and 0.5 m was also recorded. Errors that were not numerical or infinitely large were excluded before statistical analysis, and the number of invalid errors was recorded.

[0100] The two-dimensional bidirectional state space module is responsible for establishing long-distance correlations in the angle and time delay dimensions, while the time delay feature enhancement module is responsible for extracting local changes related to propagation time delay. Together with the convolutional encoder-decoder, they complete the mapping from channel state information to the angle-of-arrival-time-of-flight joint spectrum.

[0101] The final experimental results show that, regarding the joint spectrum recovery results, the original spatial root mean square error of the DenseED model is 54.918, and the peak position error is 2.902 px. After using the joint spectrum recovery network described in this invention, the original spatial root mean square error is reduced to 8.1394, and the peak position error is reduced to 0.995 px. These results indicate that combining the time delay feature enhancement module with the two-dimensional bidirectional state space module can reduce the overall error between the recovered joint spectrum and the true joint spectrum, and improve the accuracy of peak position recovery. The relevant experimental results are consistent with the comparative data in the paper. Regarding the target position calculation results, the average positioning error of the DenseED model is 0.078 m, and the maximum positioning error is 2.595 m. After using the method of this invention, the average positioning error is 0.063 m, the maximum positioning error is reduced to 0.175 m, and the number of samples with errors greater than 0.2 m and 0.5 m is zero. Therefore, the method of this invention can reduce large positioning errors and abnormal positioning results while maintaining average positioning accuracy, thus improving the stability of target position estimation.

[0102] It should be noted that the input size, number of network layers, number of candidate peaks, angle search range, flight time range, transmitter-receiver spacing, and training parameters given in the embodiments are all specific parameter settings. Those skilled in the art can adjust the relevant parameters according to the device bandwidth, number of receiving antennas, environmental size, and positioning accuracy requirements.

[0103] Those skilled in the art will understand that, without changing the basic technical concept of the present invention, the number of encoder and decoder layers, convolution kernel size, state space modeling form, delay feature enhancement branch, gating fusion method, joint loss weight, and path selection conditions can be adjusted or equivalently replaced. The scope of protection of the present invention shall be determined by the scope defined in the claims.

Claims

1. A joint spectrum recovery and target localization method based on Wi-Fi channel state information, characterized in that, Includes the following steps: Step 1: Obtain channel state information at different receiving antennas and different frequency sampling points at the receiving end, organize and preprocess the channel state information to obtain network input features; Step 2: Standardize the network input features based on the statistical parameters obtained from the input features of the training set; Step 3: The standardized network input features are fed into the joint spectrum recovery network to obtain the angle-of-arrival-time-of-flight joint spectrum; the joint spectrum recovery network is based on a convolutional encoder-decoder and is equipped with a two-dimensional bidirectional state space module and a time delay feature enhancement module; the two-dimensional bidirectional state space module performs sequence modeling along the row and column directions of the two-dimensional features respectively, and the time delay feature enhancement module extracts local change features related to propagation delay; Step 4: Perform peak search on the angle-of-arrival-time-of-flight joint spectrum, determine the target-related propagation path from the candidate spectrum peaks, and obtain the angle of arrival and time of flight corresponding to the target-related propagation path; Step 5: Based on the angle of arrival, flight time, known distance between the transmitter and receiver, and pre-established spatial geometric relationships, obtain the position of the target to be located.

2. The joint spectrum recovery and target localization method based on Wi-Fi channel state information according to claim 1, characterized in that, The data processing in steps 1 and 2 includes: The channel state information is arranged according to the receiving antenna dimension and the frequency sampling dimension to form input features suitable for the processing of the joint spectrum recovery network; Calculate the mean and standard deviation using the input features from the training set, and standardize them according to the following formula: Where x represents the input features before standardization. This represents the standardized input features. and Let represent the mean and standard deviation of the input features in the training set, respectively. A constant set to prevent the denominator from being zero; During the testing or actual localization phase, the mean and standard deviation obtained during the training phase are used to standardize the network input features.

3. The joint spectrum recovery and target localization method based on Wi-Fi channel state information according to claim 1, characterized in that, The convolutional encoder-decoder in step 3 includes an input mapping unit, a three-level coding unit, a bottleneck feature processing unit, a decoding unit corresponding to each coding unit, and a joint spectrum output unit. The input mapping unit is used to perform preliminary convolution mapping on the standardized input features; the encoding unit is used to extract local features step by step and reduce the spatial size of the feature map; the bottleneck feature processing unit is used to fuse and process the encoded deep features; the decoding unit is used to restore the feature map size step by step and concatenate the decoded features with the shallow features output by the corresponding encoding unit; the joint spectrum output unit is used to generate the angle-of-arrival-time-of-flight joint spectrum of a single channel.

4. The joint spectrum recovery and target localization method based on Wi-Fi channel state information according to claim 1, characterized in that, The two-dimensional bidirectional state space module in step 3 includes: The input two-dimensional features are organized into multiple feature sequences along the row direction, and each row feature sequence is processed using a first state space unit to obtain row-direction features. The two-dimensional features are then organized into multiple feature sequences along the column direction, and each column feature sequence is processed using a second state space unit to obtain column-direction features. The row-direction features and column-direction features are restored to a two-dimensional feature form and fused through feature concatenation and convolutional mapping to establish long-distance associations between different regions in the two-dimensional feature map.

5. The joint spectrum recovery and target localization method based on Wi-Fi channel state information according to claim 1, characterized in that, The latency feature enhancement module in step 3 includes a first convolutional branch, a second convolutional branch, a latency-aware branch, and a gated fusion unit: The first convolutional branch performs convolution processing along the frequency sampling direction related to the time of flight to extract local patterns of channel features changing with frequency; the second convolutional branch performs convolution processing along a direction orthogonal to the frequency sampling direction to preserve local structural information in the other direction; the delay-aware branch performs frequency domain transformation on the input features along the frequency sampling direction, obtains response features related to propagation delay based on the frequency domain transformation result, and adjusts the response features to the same size as the input features; the gated fusion unit generates gate weights based on the outputs of each branch, fuses the outputs of the first convolutional branch, the second convolutional branch, and the delay-aware branch, and performs residual combination with the input of the delay feature enhancement module.

6. The joint spectrum recovery and target localization method based on Wi-Fi channel state information according to claim 1, characterized in that, The training process of the joint spectral recovery network in step 3 includes: A logarithmic transformation is performed on the original angle-of-arrival-time-of-flight joint spectral labels to obtain logarithmic space labels: Where y represents the original joint spectrum label, The joint spectral label is represented after logarithmic transformation. A joint loss function is used to optimize the network parameters. This joint loss function includes logarithmic spatial mean square error, original spatial mean square error, original spatial robust loss, peak region weighted error, and peak coordinate error. The logarithmic spatial mean square error is used to constrain the overall distribution of the joint spectrum. The original spatial mean square error is used to reduce the pixel-by-pixel error of the restored joint spectrum. The original spatial robust loss is used to reduce the impact of abnormal spectral values ​​on the training process. The original spatial robust loss is preferably Charbonnier loss. The peak region weighted error is used to enhance the recovery of the region near the target spectral peak. The peak coordinate error is used to constrain the positional deviation between the predicted spectral peak and the true spectral peak.

7. The joint spectrum recovery and target localization method based on Wi-Fi channel state information according to claim 1, characterized in that, The peak search in step 4 includes: Based on the preset angle-of-arrival search grid and time-of-flight search grid, a peak search is performed on the angle-of-arrival-time-of-flight joint spectrum to obtain a preset number of candidate spectral peaks, and the spectral value, angle of arrival, and time of flight corresponding to each candidate spectral peak are recorded. The candidate spectral peaks are sorted from largest to smallest according to their corresponding spectral values. Among the first preset number of candidate spectral peaks after sorting, the candidate spectral peaks whose angle of arrival is consistent with the preset direct path angle of arrival and whose time of flight is not greater than the time delay tolerance are determined as direct path spectral peaks. From the candidate spectral peaks other than the direct path spectral peaks, the candidate spectral peaks whose angle of arrival is within a preset angle range, whose angle of arrival is different from the preset direct path angle of arrival, and whose time of flight is greater than the time delay tolerance are selected according to the sorting order, and these are taken as target-related propagation path spectral peaks. The angle of arrival and time of flight corresponding to the target-related propagation path spectral peaks are used as target position calculation parameters.

8. The joint spectrum recovery and target localization method based on Wi-Fi channel state information according to claim 1, characterized in that, Step 5, obtaining the location of the target to be located, includes: The process involves obtaining the angle of arrival and time of flight corresponding to the propagation path of the target, and using these parameters to calculate the target position; obtaining the known distance between the transmitter and receiver, and establishing the spatial geometric relationship between the transmitter, receiver, and the target to be located based on the known distance; inputting the angle of arrival, time of flight, and known distance between the transmitter and receiver into a pre-established coordinate transformation relationship to calculate the x-coordinate and y-coordinate of the target in a preset two-dimensional coordinate system; and combining the x-coordinate and y-coordinate to obtain the position coordinates of the target.

Images