A non-cooperative multi-user localization method for near-field scattering environments

Abstract

This invention discloses a non-cooperative multi-user localization method in near-field scattering environments, comprising two stages: signal source acquisition and user identification. When the number of users and the transmitted pilot signals are unknown, the method leverages the system's inherent characteristics and machine learning algorithms to achieve multi-user localization in complex scattering environments. In the first stage, utilizing the spatial orthogonality between signals from different locations in the near field, a synchronous orthogonal matching pursuit algorithm based on a deep convolutional neural network (CNN) is used, with a stop iteration indication provided, to acquire the positions of all users and scattering points. In the second stage, a signal energy-assisted k-means clustering algorithm is used, simultaneously incorporating the distance between data points and signal energy information into the cluster center update criterion to accurately identify users. This invention effectively solves the problem of non-cooperative multi-user localization in near-field scattering environments.

Description

Technical Field

[0001] This invention relates to the field of near-field user positioning technology, specifically a non-cooperative multi-user positioning method under near-field scattering environment. Background Technology

[0002] User positioning is an important research direction for sixth-generation mobile communication networks. Although GPS, based on the Global Positioning System, is widely used for navigation and positioning, GPS signals suffer severe attenuation in densely populated urban areas or indoor environments. Furthermore, the positioning accuracy of GPS in scattering environments needs improvement. In recent years, positioning technologies based on wireless local area networks have received widespread attention.

[0003] In the far field, wireless positioning typically relies on three types of measurements: Received Signal Strength (RSS), Time of Arrival (ToA), and Angle of Arrival (AoA). RSS-based methods generally perform worse than the other two, while ToA measurements implicitly require transceiver synchronization, and AoA measurements become ineffective in scattering environments. With the increase in the number of antennas and carrier frequencies, in 6G mobile communication systems, most users are likely located in the near field. In the near field, the wavefront of electromagnetic wave propagation is modeled as a spherical wave. The properties of the spherical wavefront offer new opportunities for user positioning: the user's antenna position can be inferred from the phase difference of the spherical waves received on different antennas, without requiring transceiver time synchronization. Therefore, in the near field, the curvature of arrival (CoA) of the spherical wave can be used to infer the user's location.

[0004] Near-field positioning technology has recently become a research hotspot. Many studies have presented theoretical Clumello bounds for near-field positioning accuracy, and numerous studies have been conducted on specific positioning algorithms, including maximum likelihood methods and subspace methods utilizing second-order statistics (SOS). Unlike plane wavefronts, which only contain angular information, spherical wavefronts contain both user angle and distance parameters. Therefore, extended two-dimensional versions of subspace positioning methods, such as Multi-Signal Classification (MUSIC) and Rotation Invariant Techniques (ESPRIT), have been proposed and used to obtain near-field user positions. However, the positioning accuracy of statistics-based methods depends on the estimation accuracy of the autocorrelation matrix of the received signal, and eigenvalue decomposition of the autocorrelation matrix also leads to high computational complexity. Some studies have proposed subspace positioning methods that do not require eigenvalue decomposition; however, most of these subspace-based algorithms require symmetric uniform linear arrays to utilize the Toeplitz property of the autocorrelation matrix, and their positioning performance deteriorates significantly in the presence of unknown noise or multipath fading.

[0005] On the other hand, some studies have proposed near-field localization and trajectory prediction methods based on deep learning. For example, convolutional neural networks (CNNs) are used to handle the classification and localization of mixed near-field and far-field sources. This method treats position parameter estimation as a regression problem, resulting in significantly lower computational complexity compared to methods that treat position parameter estimation as a classification problem. Other research addresses near-field user localization in the presence of broadband beam splitting effects and spatial non-stationary characteristics, designing a CNN-based localization scheme. This scheme utilizes the input and output of a controllable beam splitting scheme to generate estimates of position parameters such as angle and distance. When the direct path of the user signal to the receiver is blocked, intelligent reflector (RIS) technology offers a possibility for user localization. RIS can provide a controllable refraction path between the user and the base station. Therefore, RIS-based near-field localization technology has been extensively studied, including theoretical analyses of RIS localization in mixed near-field and far-field scenarios, as well as optimization of joint RIS parameters and design of localization algorithms. However, RIS-based localization suffers from synchronization issues among the user, RIS, and base station. Summary of the Invention

[0006] To overcome the problem of non-cooperative near-field multi-user localization under conditions of unknown user numbers and pilot signal transmission, this invention provides a non-cooperative multi-user localization method in near-field scattering environments. This method fully utilizes the system's own characteristics and machine learning algorithms to achieve multi-user localization in complex scattering environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a non-cooperative multi-user localization method in a near-field scattering environment, comprising a new algorithm framework, which includes two stages: signal source acquisition and user identification. When the number of users is unknown and pilot signals are transmitted, multi-user localization is achieved in a complex scattering environment by leveraging the system's own characteristics and machine learning algorithms.

[0008] The method includes the following steps:

[0009] 1) In the first stage, the location of all users and scattering points is obtained using the Synchronous Orthogonal Matching Pursuit (SOMP) algorithm based on a deep convolutional neural network (CNN).

[0010] 2) In the second stage, the signal energy-assisted k-means clustering algorithm is used to incorporate both the distance between data points and the signal energy information into the update criteria of the cluster centers, thereby accurately identifying users.

[0011] Preferably, in step 1), the Synchronous Orthogonal Matching Pursuit (SOMP) algorithm fully utilizes the sparsity of the channel and the orthogonality between signals from different sources in the near-field multi-user localization problem to achieve location acquisition under unknown user pilot signals. This algorithm is based on the following sparsified representation model:

[0012] Y = DU + W

[0013] Where Y = [y1,...,y T ] represents the received signal matrix at T time points, and the received signal vector y at each time point. t The received signal contains M antennas; similarly, the matrix U = [u1,...,u T It consists of sparse vectors at T time points, each sparse vector u t Received signal t t The projection onto the dictionary space, matrix W = [w1,...,w T [] represents noise; due to the sparsity of the channel, the rows of matrix U are sparse; therefore, the localization problem can be solved using the SOMP algorithm; in addition, D in the above formula represents the near-field dictionary matrix, which constitutes a complete angle-range space composed of the guidance vector b(θ,r), that is, the dictionary covers the guidance vectors under all possible angles and ranges; specifically, the angle space is divided into θ i , i = 1,...,M, where cosθ i ∈[-1,1]; divide the distance space into r j =[r min ,r max If j = 1, ..., N, then dictionary D is represented as:

[0014] D=[b(θ1,r1),...,b(θ1,r N ),...,b(θ M ,r1),...,b(θ M ,r N )]

[0015] in,

[0016]

[0017] The above equations are the expressions for the guidance vector and the distance from the data source (θ,r) to the m-th antenna, respectively, where fc represents the carrier frequency, c is the speed of light, and d represents the antenna spacing, set to half a wavelength.

[0018] The sparse vector u obtained after projecting the received signal onto the dictionary D is... t It is considered as an indicator vector with MN elements, each of which is either zero or non-zero; a non-zero value indicates the presence of a user or scatterer in the corresponding grid of the dictionary column space.

[0019] Preferably, in step 1), the deep convolutional neural network (CNN) provides a stop iteration indication for the SOMP algorithm. Specifically, the CNN determines whether the residual signal after each iteration is close to the ambient noise. If not, the SOMP iteration will continue, and another path will be extracted from the received signal; if so, the algorithm terminates.

[0020] Preferably, in step 1), the constructed deep convolutional neural network (CNN) structure includes two CBR fusion blocks, a pooling layer, several fully connected (FC) layers, and a softmax layer, wherein the CBR fusion block contains a convolutional layer, a batch normalization layer, and a rectified linear unit (ReLU) layer.

[0021] Preferably, in step 2), the signal energy-assisted k-means clustering algorithm iteratively finds all users and their surrounding scattering points; each iteration includes three steps: distance calculation, data point allocation, and cluster center update; in the first step, the distance from all data points to each cluster center is calculated using Euclidean distance, i.e., the distance from data point s to each cluster center is 1. i to cluster center s' j The distance is expressed as:

[0022]

[0023] Among them, s i Vector v represents the i-th data point among a large number of data points. i =[x i ,y i ] T Representing point s i polar coordinates (θ) i ,r i The corresponding Cartesian coordinates.

[0024] Preferably, in step 2), after the data points are allocated, the maximum signal energy value in each cluster is detected. When the maximum value is lower than the average energy of all input data points, it is considered that there is a high probability that there are no users in the cluster. Therefore, the total number of clusters is reduced by 1, and the data points of the cluster will be integrated into other clusters in the next iteration.

[0025] Preferably, in step 2), the update of cluster centers incorporates two parameters: signal energy and distance, as a metric.

[0026]

[0027] Among them, (x k' ,y k' ) is polar coordinate (θ) k' ,r k') represents the corresponding Cartesian coordinates, and k' is an index variable used to traverse the cluster C. k Each data point in the data;

[0028] ρ k' Defined as:

[0029]

[0030] Where, ρ k' Indicates cluster c k The magnitude of each data point |u k' Weighting coefficients.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] 1. Solve the positioning problem in complex scattering environments.

[0033] The method proposed in this invention can effectively solve the multi-user positioning problem in near-field scattering environments when the number of users and the transmitted pilot signal are unknown, breaking through the limitations of traditional methods in complex environments; moreover, this method is applicable to a variety of complex scattering scenarios, such as dense urban environments and indoor scenarios, and has wide applicability.

[0034] 2. Innovation in the signal source acquisition stage

[0035] This invention proposes a Synchronous Orthogonal Matching Pursuit (SOMP) algorithm based on a deep convolutional neural network (CNN). When the number of users and scatterers is unknown, the introduction of CNN provides a stopping indication for the SOMP algorithm, solving the problem that the traditional SOMP algorithm has difficulty in determining the stopping condition for iteration. Moreover, the introduction of CNN enhances the accuracy of signal source acquisition, especially in complex scattering environments, enabling more accurate location of users and scattering points.

[0036] 3. Innovation in the user identification stage

[0037] This invention proposes a signal energy-assisted k-means clustering algorithm that incorporates both the distance between data points and signal energy information into the update criterion for cluster centers, overcoming the limitation of traditional k-means algorithms that only cluster based on distance. Furthermore, by combining signal energy information, it significantly improves the accuracy of user identification, especially in near-field scattering environments, enabling more accurate user identification.

[0038] 4. Overall advantages of the algorithm framework

[0039] The algorithm framework proposed in this invention includes signal source acquisition and user identification. While ensuring positioning accuracy, it improves computational efficiency, is suitable for real-time positioning needs, and exhibits strong robustness in complex scattering environments, enabling it to stably complete multi-user positioning tasks.

[0040] 5. Practical application value

[0041] This invention can be applied to wireless communication systems, especially in dense urban environments or indoor scenarios, where it can effectively locate multiple users and improve the resource allocation efficiency of the communication system. In Internet of Things (IoT) applications, this method can help locate a large number of unknown IoT devices, improving network management and device monitoring capabilities. Meanwhile, in emergency rescue and security monitoring scenarios, this method can quickly locate multiple targets, improving rescue efficiency and monitoring accuracy. Attached Figure Description

[0042] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0043] In the attached diagram:

[0044] Figure 1 This is a schematic diagram of a multi-user system under near-field scattering conditions;

[0045] Figure 2 This is a schematic diagram of the multi-user positioning algorithm framework in this invention;

[0046] Figure 3 This is a schematic diagram of the CNN network structure in this invention;

[0047] Figure 4 This is a graph showing the training performance of the CNN network in this invention;

[0048] Figure 5 This is a test performance graph of the CNN network in this invention;

[0049] Figure 6 This is a diagram illustrating the clustering effect of the signal energy-assisted k-means algorithm in this invention;

[0050] Figure 7 This is a clustering performance graph of the signal energy-assisted k-means algorithm in this invention;

[0051] Figure 8 This is a positioning accuracy diagram of the multi-user positioning algorithm in this invention. Detailed Implementation

[0052] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0053] Example: This invention provides a non-cooperative multi-user localization method in a near-field scattering environment, comprising the following steps:

[0054] like Figure 1 As shown, consider a single-cell massive MIMO system where the base station (BS) is equipped with M antennas in the form of a uniform linear array (ULA). Typically, in 6G networks, M is assumed to be hundreds to thousands. Deploying such a large antenna array, coupled with the use of high-frequency carriers, means that the near-field region can extend to tens to hundreds of meters. Figure 1 As shown, the near-field region considered in this embodiment is located between the Fresnel boundary and the Fraunhofer boundary. The former is the outer boundary of the induced near-field, and the latter is the boundary between the radiating near-field region and the far-field region, defined respectively as:

[0055] and

[0056] Where D is the antenna diameter and λ is the carrier wavelength.

[0057] Assume there are K single-antenna users and some scatterers randomly distributed in the near-field region of BS, and the electromagnetic wave wavefront is a spherical wave. The first antenna of the uniform linear array ULA is used as the reference antenna, and the antenna spacing is... Assume the ULA is parallel to the y-axis, and the coordinates of its m-th antenna are (0, (m-1)d), where m = 1, ..., M. In this system, the BS array guidance vector corresponding to a signal at a distance r and an incident angle θ is:

[0058]

[0059] The formula for calculating the distance from the signal source to the m-th antenna is:

[0060]

[0061] For the k-th user, assume that besides the direct path, there are L surrounding paths. k If there are scatterers, then their channel vectors are expressed as:

[0062]

[0063] In this equation, the first term on the right-hand side represents the Loss of Sorrow (LoS) path, and the second term is the L path. k The sum of scattering paths. Symbol g l,k θ l,k and r l,k Let represent the coefficient, arrival angle, and distance of the l-th path, respectively. And ψ l,k The phase shift α is caused by the l-th scatterer. l,k This is the distance between the user and the l-th scatterer. Assume there are K users in total, and let s denote the signal emitted by the k-th user as s. k The signal received at BS is:

[0064]

[0065] Here, vector w represents the noise at the receiver.

[0066] like Figure 2 As shown, in order to obtain multi-user location information from the received signal, the received signal is first reformulated as follows:

[0067]

[0068] Where L = K + L1 + L2 + ... + L k It is the total number of paths through which the signal is superimposed at the receiving end, including the direct path (LoS) for each user and the refracted paths (NLoS) of all scatterers. s represents the equivalent coefficient of each path. l Let y represent the signal transmitted via path l. From the above equation, it can be seen that the signal received by the base station is the superposition of signals from all paths. Therefore, information about all paths, i.e., the angle of arrival θ of the signal source on each path, can be extracted from y. l and the distance r to BS l Parameters such as these.

[0069] To obtain these parameters, considering the sparsity of the near-field channel (i.e., the number of paths is finite), this embodiment proposes a compressed sensing (CS)-based method. This method utilizes the approximate orthogonality of the pilot vectors of different paths to obtain the signal source's location information. Unlike subspace-based methods, the CS-based method does not require calculating the autocorrelation matrix of the received signal, which reduces the computational complexity of pilot overhead and matrix eigenvalues. To utilize the CS-based method, we re-represent the aforementioned received signal y as follows:

[0070] y = Du + w

[0071] Here, D represents the near-field dictionary matrix, which consists of a complete angle-range space formed by the guidance vector b(θ,r). That is, this dictionary covers the guidance vectors for all possible angles and ranges. Specifically, the angle space is divided into θ... i , i = 1,...,M, where cosθ i ∈[-1,1]; divide the distance space into r j =[r min ,r max Let j = 1, ..., N. Then the dictionary D is represented as:

[0072] D=[b(θ1,r1),...,b(θ1,r N ),...,b(θ M ,r1),...,b(θ M ,rN )]

[0073] In the above equation, vector u is a sparse vector representing the projection of the received signal into the dictionary space. It can actually be viewed as an indicator vector with MN elements, each element being either zero or non-zero. A non-zero element indicates that there is a signal source on the corresponding grid in the column space of the dictionary.

[0074] The current problem is to obtain the positions of the non-zero elements in u, and then obtain the path parameters by indexing the relevant columns in the dictionary D. This example solves this problem using the Orthogonal Matching Pursuit (OMP) algorithm. This embodiment assumes there are T received signals at different times, thus obtaining a multiple measurement vector (MMV) version of the received signal:

[0075] Y = DU + W

[0076] Where Y = [y1,...,y] T ],U=[u1,...,u T ], and W = [w1,...,w T Let U represent the received signal matrix, sparse vector matrix, and noise matrix at T time points, respectively. Due to the sparsity of the channel, matrix U is row-sparse. Therefore, the localization problem can be solved using SOMP, a simultaneous version of OMP.

[0077] like Figure 3 As shown, since the number of users and scatterers is unknown, this invention considers using a deep convolutional network (CNN) to provide an iterative stopping indicator for the SOMP algorithm.

[0078] Specifically, the network consists of two CBR fusion blocks and several fully connected (FC) layers. The CBR fusion blocks include convolutional layers, batch normalization (BN) layers, and rectified linear unit (ReLU) layers. Since the received signal is complex, the number of input neurons is set to twice the number of receiving antennas M to process the real and imaginary parts of the received signal. In this embodiment, by dividing the signal received on each antenna into real and imaginary parts, the neural network input data at each time step is represented as follows:

[0079]

[0080] In the formula z t,m Let m represent the remaining signal on the m-th antenna, where m is the signal on the m-th antenna. and These represent the real part and the imaginary part, respectively. Starting from the sequence input layer, the input signal... The data is passed to the CBR fusion layer: First, a one-dimensional convolution is used. By sliding a kernel function or filter across the input data and calculating the dot product between the input data and the kernel, the spatial correlation of the input features is extracted. The mathematical expression is:

[0081]

[0082] Among them, w conv This represents the weights of the convolutional layer. The padding operation adjusts z... t,conv and They have the same size. And when F filters are used, the output signal comprises a total of F channels, denoted as: After convolution, a batch normalization (BN) operation is performed to obtain:

[0083]

[0084] in and σ 2 (z t,conv ) are respectively z t,conv The mean and variance are given. The ∈ sign is added to the formula for numerical stability, and:

[0085]

[0086]

[0087] This layer normalizes the input to have zero mean and unit variance. This operation reduces the vanishing gradient problem and improves the network's robustness to hyperparameters.

[0088] Subsequently, a ReLU layer is applied to the output of the BN layer, i.e., the activation function f(x) = max(0,x), to obtain the z-value of each channel. outTo facilitate the network's learning of nonlinear relationships, the ReLU operation maps negative input values ​​to zero while keeping positive input values ​​unchanged. After the two CBR fusion blocks, a max-pooling layer is used for downsampling. Its output is reconstructed and passed to a fully connected (FC) layer. In this network, this embodiment uses a 2M×100 FC layer, a 100×200 FC layer, and a 200×2 FC layer, with ReLU activation functions interspersed between the FC layers. From a representation learning perspective, convolutional and pooling layers in a convolutional neural network can extract features from the input data, while the FC layer's role is to combine the extracted features in a nonlinear manner to obtain the output. The FC layer itself does not have feature extraction capabilities but instead attempts to use existing higher-order features to achieve the learning objective. Finally, the output of the last FC layer is processed by a softmax layer to calculate the probability of each class, i.e., whether the remaining signal still contains a valid path or is close to environmental noise. This embodiment has a total of T signals at different times. Each signal is input into the network individually to obtain an output. The final result is obtained through a voting method. That is, when more than T / 2 outputs indicate that the remaining signals contain a valid path, this embodiment will continue to iterate the SOMP algorithm.

[0089] Figure 4 and Figure 5 The training and testing performance graphs of the CNN network in this invention are shown in the figures. Figure 4 The changes in accuracy and loss over iterations are shown. The network was trained using 5000 training samples, minimizing cross-entropy loss using the Adam optimizer, with a learning rate of 0.03, 100 epochs, and a mini-batch size of 20 observations. The training samples consisted of two classes: useful signals from the user or scattering object, and white noise. These two signals were randomly generated and labeled as 1 and 0, respectively, and normalized before being input into the network. Due to the small number of training samples and the simple architecture of the CNN, training on a standard CPU only takes a few minutes. Figure 4 The training results demonstrate superior performance, exhibiting high accuracy and low loss. Results on the test samples further validate its generalization ability, achieving a classification error rate of 0.0027 on 3000 samples. Figure 5 As shown, the left figure is the label distribution of the test data, and the right figure is the prediction result.

[0090] Figure 6This paper demonstrates the clustering effect of the signal energy-assisted k-means clustering algorithm presented in this invention. The algorithm attempts to find all users and their surrounding scatterers through iterative clustering. Each iteration includes three steps: calculating the distance from all data points to the cluster centers, point assignment, and updating the cluster centers. More specifically, in the distance calculation, this embodiment calculates the distance between L input data points and each cluster center. In most applications, the distance metric is Euclidean distance, i.e., the distance between data points s... i To the cluster center s' j The distance between them is expressed as:

[0091]

[0092] Among them, s i Vector v represents the i-th data point among a large number of data points. i =[x i ,y i ] T Representing point s i polar coordinates (θ) i ,r i The corresponding Cartesian coordinates.

[0093] In the point assignment, based on the distance calculated in the previous step, each data point in set S is assigned to the nearest cluster, resulting in k clusters c1,...,c k .

[0094] The cluster center update step here serves two purposes: adjusting the number of clusters and updating the centers of each cluster. For each cluster, if the maximum signal energy value of its data points is still lower than the given average signal energy value, it is considered that there are no users in that cluster, and the data points of that cluster will be merged into other clusters in the next iteration. This scenario is reasonable in this embodiment because the signal from users is usually much stronger than the signal from scatterers. Therefore, when the maximum value in a cluster is still relatively small, it is very likely that there are no users in that cluster, or the signal of that user is very weak and can be ignored. Then, for other clusters, this embodiment proposes an amplitude-weighted cluster center update criterion, namely:

[0095]

[0096] Among them, (x k' ,y k' ) is polar coordinate (θ) k' ,r k' ) represents the corresponding Cartesian coordinates, ρ k' Indicates cluster c k The amplitude of each data point |u k' The weighting coefficients are defined as follows:

[0097]

[0098] Compared to the traditional k-means clustering algorithm, which only uses location to calculate cluster centers, the metric proposed in this invention involves both the amplitude and location information of each data point within a cluster. This allows the embodiment to fully utilize prior knowledge of the considered scenario: each user and their associated scatterers form a relatively small cluster, and the signal transmitted through the LoS path is typically stronger than the signal from the scatterers, thus pushing the cluster centers closer to the users within each cluster. Therefore, with algorithm iteration, user-centric clustering will be formed.

[0099] This process continues until the number of clusters and the location of each cluster center no longer change significantly. Then, the cluster center is replaced with the point with the largest amplitude in that cluster, which is the finally identified user.

[0100] Figure 7 This is a clustering performance graph of the energy-assisted k-means clustering algorithm in this invention. Figure 7 In (a), accuracy is defined. To evaluate the clustering accuracy of the proposed method, where n(c k Let be the number of data points obtained after clustering user k, and n be the total number of data points. In contrast, [the remaining text appears to be incomplete and requires further context.] Figure 7 In (b), the definition is... To evaluate the effectiveness of the proposed method in identifying users, among which K represents the number of users identified, and K represents the total number of users. Higher values ​​for both metrics indicate better clustering performance. Figure 7 As shown in (a), the proposed unsupervised clustering algorithm exhibits excellent performance, achieving an accuracy of over 96% in scenarios with varying numbers of users. However, the clustering accuracy decreases slightly with increasing user numbers. This is because, with a larger user base, the scatterers generated in the simulation setup may overlap or be very close in space, leading to clustering errors. However, this loss is negligible. Figure 7 As can be seen in (b), the user identification accuracy of the algorithm is close to 100%. Since user identification is the main task of multi-user localization, it can be concluded that the method has achieved excellent performance in non-cooperative multi-user localization.

[0101] Figure 8 Simulation curves showing the algorithm's positioning accuracy as a function of the number of users are presented. Positioning accuracy is represented by the mean square error of the location estimation results. In the simulation, the base station used 256 antennas at a carrier frequency of 28 GHz, utilizing received signals at 10 time points. Users were randomly distributed within a near-field range of 10-50 meters. Figure 8As can be seen, the positioning accuracy can reach an error of about 0.8 meters under different numbers of users, indicating that the proposed algorithm effectively solves the non-cooperative multi-user positioning problem in near-field scattering environment.

[0102] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-cooperative multi-user localization method under near-field scattering environment, characterized in that: It includes two stages: signal source acquisition and user identification. When the number of users is unknown and pilot signals are being transmitted, multi-user positioning is achieved in complex scattering environments by leveraging the system's own characteristics and machine learning algorithms. The method includes the following steps: 1) In the first stage, the synchronous orthogonal matching tracking algorithm based on deep convolutional neural network (CNN) is used to obtain the positions of all users and scattering points; 2) In the second stage, the signal energy-assisted k-means clustering algorithm is used to incorporate both the distance between data points and the signal energy information into the update criteria of the cluster centers, thereby accurately identifying users; In step 1), the synchronous orthogonal matching pursuit algorithm is based on the following sparse representation model: ； in, express The received signal matrix at each time step, and the received signal vector at each time step. Include The received signal from the root antenna; similarly, the matrix Depend on The sparse vectors at each time step are constructed, and each sparse vector... Received signal Projection in dictionary space, matrix This represents noise; additionally, in the above formula, D represents the near-field dictionary matrix, derived from the guiding vector. A complete angle-distance space is constructed, meaning the dictionary covers the guide vectors for all possible angles and distances; this is achieved by dividing the angle space into... , ,in, Divide the distance space into , Then dictionary D is represented as: ； in, ； ； The above equations represent the expression for the guidance vector and the data source, respectively. To the The distance between the antennas, where Indicates the carrier frequency. At the speed of light, This indicates the antenna spacing, set to half a wavelength. The sparse vector obtained after projecting the received signal onto the dictionary D is... Considered a person with An indicator vector of elements; In step 2), during the update of cluster centers, two parameters, signal energy and distance, are introduced as metrics, namely: ， ； in, polar coordinates Here, k' represents the corresponding Cartesian coordinates, and k' is an index variable used to traverse the cluster C. k Each data point in the data; Defined as: ； in, Cluster amplitude of each data point Weighting coefficients.

2. The non-cooperative multi-user positioning method under near-field scattering environment according to claim 1, characterized in that: In step 1), the deep convolutional neural network (CNN) provides a stop iteration indication for the SOMP algorithm. Specifically, the CNN determines whether the residual signal after each iteration is close to the ambient noise. If not, the SOMP iteration will continue, and another path will be extracted from the received signal. If so, the algorithm terminates.

3. The non-cooperative multi-user positioning method under near-field scattering environment according to claim 2, characterized in that: In step 1), the constructed deep convolutional neural network (CNN) structure includes two CBR fusion blocks, a pooling layer, several fully connected layers, and a Softmax layer. The CBR fusion block contains a convolutional layer, a batch normalization layer, and a modified linear unit layer.

4. The non-cooperative multi-user positioning method under near-field scattering environment according to claim 1, characterized in that: In step 2), the signal energy-assisted k-means clustering algorithm iteratively finds all users and their surrounding scattering points. Each iteration includes three steps: distance calculation, data point allocation, and cluster center update. In the first step, the distance from all data points to each cluster center is calculated using Euclidean distance, i.e., the distance from each data point to the cluster center. To the cluster center The distance is expressed as: ； Among them, s i A vector representing the i-th data point among many data points. Point polar coordinates The corresponding Cartesian coordinates.

5. The non-cooperative multi-user positioning method under near-field scattering environment according to claim 1, characterized in that: In step 2), after the data points are allocated, the maximum signal energy value in each cluster is detected.

Images