Autoencoder-aided de-celled massive MIMO low-dimensional fingerprinting positioning method

Abstract

The application discloses a kind of autoencoder-assisted de-cell large-scale MIMO low-dimensional fingerprint positioning method, comprising the following steps: step 1: extracting the angle information matrix of reference point user;Step 2: training autoencoder network;Step 3: fingerprint clustering based on K-means++;Step 4: define the similarity criterion between position fingerprints;Step 5: extract user fingerprint information;Step 6: fingerprint matching and position estimation;Step 7: estimate user position.The application utilizes autoencoder network to learn the key information hidden in angle domain channel matrix, extracts deep position-channel mapping relationship, and removes redundant information.Under the premise of not significantly affecting positioning accuracy, effectively reduce the dimension of position fingerprint, improve the timeliness of fingerprint positioning scheme in outdoor environment.Kmeans++ algorithm is used to cluster offline fingerprints, which further improves the online matching efficiency of fingerprint positioning without affecting the positioning accuracy.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and relates to the positioning process in fifth-generation (5G) mobile communication, and particularly to an autoencoder-assisted decellularized large-scale MIMO low-dimensional fingerprint positioning method. Background Technology

[0002] In the optimization design of 5G communication systems, user location information can assist in channel estimation, simplify beam scanning, and optimize initial access, thus helping to further improve the performance of the communication system. The 3GPP Rel-16 release indicates that the communication network can guarantee 80% of users a positioning accuracy of 3m indoors and 10m outdoors, leading to widespread attention being paid to wireless positioning technologies based on these networks. Traditional wireless positioning technologies, represented by geometric positioning, rely on line-of-sight (LoS) propagation of signals, using signal processing methods to estimate location parameters related to the user's location, such as Received Signal Strength (RSS), Angle of Arrival (AoA), or Time-of-Arrival (ToA). However, in non-line-of-sight (NLoS) propagation environments, the direct path is obstructed, and the location parameters estimated by geometric positioning schemes map to reflection path parameters, severely impacting their positioning accuracy.

[0003] To improve the positioning accuracy of wireless networks in scattering-rich environments, fingerprint positioning technology based on channel feature matching has developed rapidly. Due to the influence of user location, the multipath scattering components of a signal exhibit different characteristics. These characteristics vary with the user's location and can serve as the user's wireless identifier, also known as a location fingerprint. Common location fingerprints mainly include RSS, AoA, and channel state information. The fingerprint positioning process is divided into offline and online stages: In the offline stage, the base station collects pilot signals transmitted by users at a designated location (reference point) within the service area and extracts location fingerprints to construct an offline fingerprint database; in the online stage, online fingerprints are extracted from the pilot signals of the user to be located and matched with the offline fingerprints of the reference points. The fingerprint similarity is ranked to find nearest neighbor reference points. Finally, the known location information of the nearest neighbor reference points is used to estimate the user's location.

[0004] To improve the spectral efficiency of communication networks, network densification technologies, such as cell splitting, have continuously evolved, leading to a continuous reduction in cell radius. This has made inter-cell interference a bottleneck restricting system performance. To overcome this bottleneck, decellularized massive MIMO (Multi-Input Multi-Output) systems have gained widespread industry acceptance. A large number of access points (APs) are deployed dispersedly within the service area, utilizing the same time-frequency resources to provide services to users, thus eliminating inter-cell interference. The closer distance between users and APs significantly improves macro-diversity gain and further enhances system spectral efficiency, making it considered one of the potential key technologies for 5G evolution and even 6G. Decellularized massive MIMO provides strong support for wireless positioning: 1. Distributed APs can acquire richer location information and are more robust in complex environments; 2. The closer distance between APs and users increases the probability of Loss of Position (LoS) links, simplifying positioning scheme design and improving positioning accuracy; 3. Massive MIMO effectively improves the system's spatial resolution. Research on positioning technology in decellularized massive MIMO systems has become a very important research direction.

[0005] Currently, some literature has conducted relevant research on fingerprint positioning in massive MIMO systems. For example, the literature "Qiu J, Xu K, Shen Z. Cooperative fingerprint positioning for cell-free massive MIMO systems[C] / / 2020 International Conference on Wireless Communications and Signal Processing(WCSP).IEEE,2020:382-387." proposes a positioning method that combines RSS and AoA fingerprints. The basic idea is to extract the user's angular domain channel matrix and use its Hadamard product as the location fingerprint to achieve user positioning. This method has similar applications in the literature “Qiu J, Xu K, Xia X, et al. Secure transmission scheme based on fingerprint positioning in cell-free massive MIMO systems[J].IEEE Transactions on Signal and Information Processing over Networks, 2022, 8: 92-105.” and “Wei C, Xu K, Shen Z, et al. Fingerprint-Based Localization and Channel Estimation Integration for Cell-Free Massive MIMO IoT Systems[J].IEEE Internet of Things Journal, 2022.”. This scheme integrates user angle of arrival information and power information, enriching the expressive power of location fingerprints, and further reducing positioning errors by combining the advantages of multiple APs in decellularized massive MIMO systems. However, the fingerprint dimensionality is high in this scheme, affecting the scalability and timeliness of positioning. In addition, increasing the array antenna aperture can significantly improve the system spatial resolution, and increasing the number of positioning APs can effectively reduce positioning errors; however, the fingerprint dimensionality also increases accordingly, seriously affecting the fingerprint positioning efficiency in outdoor environments. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, this invention provides an autoencoder-assisted decellularized large-scale MIMO low-dimensional fingerprint localization method. Its purpose is to solve the following technical problems: utilizing an autoencoder network to learn key information hidden in the angular domain channel matrix, extracting deep-level position-channel mapping relationships, and removing redundant information. This effectively reduces the dimensionality of the location fingerprint without significantly affecting localization accuracy, improving the timeliness of fingerprint localization schemes in outdoor environments. Furthermore, the K-means++ algorithm is used to cluster offline fingerprints, further improving the online fingerprint matching efficiency without affecting localization accuracy.

[0007] To achieve the above and other related objectives, the present invention provides an autoencoder-assisted decellularized large-scale MIMO low-dimensional fingerprint localization method, comprising the following steps:

[0008] Step 1: Extract the angle information matrix of the reference point user

[0009] Assume that in a decellularized massive MIMO system, N access points (APs) are randomly distributed within the service area. Each AP is equipped with a uniform linear antenna array, with M elements. The service area is uniformly divided, and the intersection of the grid points serves as reference points, denoted as R. All APs estimate the channel using uplink pilot signals to obtain the CSI at each reference point. Considering a narrowband multipath channel model, the channel between the nth AP and the rth access point (RP) is:

[0010]

[0011] Where: L is the number of resolvable scattering paths. Let β be the complex gain of the l-th path. nr The large-scale fading coefficient, comprising path loss and shadow fading, is expressed as:

[0012]

[0013] Where: d nr Let the distance be the Euclidean distance between the nth AP and the rth RP. Indicates the fading of the shadow; This represents the array response vector corresponding to the l-th scattering path.

[0014] Where d is the antenna spacing and λ is the carrier wavelength. The signal arrival angle of the l-th path;

[0015] The uplink pilot signal received at the nth AP from the rth RP is:

[0016]

[0017] Where: φ r represents the uplink pilot sequence, and w represents the additive white Gaussian noise vector;

[0018] The channel can be estimated based on the received uplink pilot signal, denoted as:

[0019] Performing a Fourier transform on the estimated channel yields the angular domain channel:

[0020]

[0021] Where: F represents the DFT matrix.

[0022] The angle-domain channel matrix is ​​represented as:

[0023]

[0024] To suppress fingerprint fluctuations caused by small-scale fading, the angle-domain channel matrix across multiple coherent time intervals is further processed as follows:

[0025]

[0026] Step 2: Train the autoencoder network

[0027] The Φ obtained in step 1 r As a training dataset, it is input into the autoencoder network to learn the intrinsic relationship between the channel information matrix and the user's location, and to extract low-dimensional fingerprint information; input data [Φ r ] i The data is sent to the input layer, passed through one or more hidden layers, and then to the output layer to obtain the output data. Automatic encoders reduce the input data [Φ r ] i With output data The error between them is used to learn the inherent features hidden in the original data, i.e., the output matrix of the hidden layer [Θ]. r ] i After the network training is complete, the hidden layer output matrix Θ will be displayed. r Fingerprint as a reference point location;

[0028] Step 3: Fingerprint clustering based on K-means++

[0029] (1) Randomly select Q reference point locations and their vectorized fingerprints as cluster centers, denoted as .

[0030] (2) Define the Euclidean distance L(Θ) r ,μ i )=||vec(Θ r)-μ i || 2 ;

[0031] (3) For each reference point location fingerprint Θ r Calculate its relationship with each cluster center μ i The Euclidean distance is used to assign the fingerprint of the reference point to the cluster corresponding to the minimum value of the Euclidean distance;

[0032] (4) For each cluster, recalculate its cluster center:

[0033]

[0034] Where t represents the t-th iteration; |c i | indicates that it belongs to cluster c i The number of reference points;

[0035] (5) Define the loss function

[0036]

[0037] Repeat steps (3) and (4) until all cluster centers are fixed, i.e., the loss function J < ε, where ε is the loss threshold.

[0038] (6) Rewrite the converged cluster centers in matrix form i dim×N , where dim is the output dimension of the autoencoder;

[0039] Step 4: Define the similarity criteria between location fingerprints

[0040] The compressed location fingerprint retains the main user location information, and the Euclidean distance between location fingerprints is defined as the fingerprint similarity criterion:

[0041]

[0042] Among them, the closer the Euclidean distance between location fingerprints, the higher the similarity between the fingerprints;

[0043] Step 5: Extract user fingerprint information

[0044] The user channel state information is estimated using the pilot signal transmitted by user k, denoted as . The angle domain channel is obtained by performing a Fourier transform. Then The compressed data is extracted from the hidden layer of the trained autoencoder and used as the user's location fingerprint, denoted as Θ. k ;

[0045] Step 6: Fingerprint matching and location estimation

[0046] First, calculate the user location fingerprint Θ according to the similarity judgment criteria obtained in step 4. k The similarity Λ(Θ) between each fingerprint cluster center k ,μ i Select the two clusters with the highest similarity, denoted as Ξ; then compare the user's fingerprint with the reference point fingerprints within Ξ one by one, and select the three nearest neighbors with the highest similarity, denoted as Ξ. The corresponding similarity coefficient is denoted as

[0047] Step 7: Estimate user location

[0048] The Weighted K-Nearest Neighbors (JNNN) algorithm, an improvement on the KNNN algorithm, achieves more accurate localization using the same number of nearest neighbors (RPs). Its formula is expressed as:

[0049]

[0050] Due to the application of the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0051] (1) By extracting the intrinsic relationship between user location and channel features through an autoencoder network, location fingerprint compression is achieved, which helps to reduce storage pressure and improve location fingerprint matching efficiency;

[0052] (2) Compared with other dimensionality reduction algorithms, autoencoder networks can accurately extract position information in the channel with little impact on positioning accuracy;

[0053] (3) Based on location fingerprint compression, the Kmean++ algorithm is used to cluster offline location fingerprints, which further improves the matching efficiency.

[0054] Instruction manual illustrations

[0055] Figure 1 This is a schematic diagram of the technical solution of the present invention;

[0056] Figure 2 This is a schematic diagram of the system model of the decellularized large-scale MIMO fingerprint positioning method of the present invention;

[0057] Figure 3 This is a schematic diagram of the network structure of the automatic encoder used in this invention;

[0058] Figure 4 This is a schematic diagram of the cumulative distribution function of localization error compared to other dimensionality reduction algorithms. Figure 1 ;

[0059] Figure 5 This is a schematic diagram of the cumulative distribution function of localization error compared to other clustering algorithms. Figure 2 . Detailed Implementation

[0060] The following describes the implementation of the present invention with reference to specific embodiments and accompanying drawings. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification.

[0061] Example:

[0062] An autoencoder-assisted decellularized large-scale MIMO low-dimensional fingerprint localization method, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0063] Step 1: Extract the angle information matrix of the reference point user

[0064] In a decellularized massive MIMO system, N access points (APs) are randomly distributed within the service area. Each AP is equipped with a uniform linear antenna array, denoted by M elements. The service area is uniformly divided, and the intersections of the grids are reference points, denoted by R. Figure 2 As shown, all access points (APs) estimate the channel using uplink pilot signals to obtain the CSI at each reference point; considering the narrowband multipath channel model, the channel between the nth AP and the rth RP is:

[0065]

[0066] Where: L is the number of resolvable scattering paths. Let β be the complex gain of the l-th path. nr The large-scale fading coefficient, comprising path loss and shadow fading, is expressed as:

[0067]

[0068] Where: d nr Let the distance be the Euclidean distance between the nth AP and the rth RP. Indicates the fading of the shadow; This represents the array response vector corresponding to the l-th scattering path.

[0069] Where d is the antenna spacing and λ is the carrier wavelength. The signal arrival angle of the l-th path;

[0070] The uplink pilot signal received at the nth AP from the rth RP is:

[0071]

[0072] Where: φ r represents the uplink pilot sequence, and w represents the additive white Gaussian noise vector;

[0073] The channel can be estimated based on the received uplink pilot signal, denoted as:

[0074] Performing a Fourier transform on the estimated channel yields the angular domain channel:

[0075]

[0076] Where: F represents the DFT matrix.

[0077] The angle-domain channel matrix is ​​represented as:

[0078]

[0079] To suppress fingerprint fluctuations caused by small-scale fading, the angle-domain channel matrix across multiple coherent time intervals is further processed as follows:

[0080]

[0081] Step 2: Train the autoencoder network

[0082] The Φ obtained in step 1 r As a training dataset, it is input into the autoencoder network to learn the intrinsic relationship between the channel information matrix and the user's location, and to extract low-dimensional fingerprint information; the network structure of the autoencoder is as follows. Figure 3 As shown, the input data [Φ r ] i The data is sent to the input layer, passed through one or more hidden layers, and then to the output layer to obtain the output data. Automatic encoders reduce the input data [Φ r ] i With output data The error between them is used to learn the inherent features hidden in the original data, i.e., the output matrix of the hidden layer [Θ]. r ] i After the network training is complete, the hidden layer output matrix Θ will be displayed. r Fingerprint as a reference point location;

[0083] Step 3: Fingerprint clustering based on K-means++

[0084] (1) Randomly select Q reference point locations and their vectorized fingerprints as cluster centers, denoted as .

[0085] (2) Define Euclidean distance

[0086] (3) For each reference point location fingerprint Θ r Calculate its relationship with each cluster center μ iThe Euclidean distance is used to assign the fingerprint of the reference point to the cluster corresponding to the minimum value of the Euclidean distance;

[0087] (4) For each cluster, recalculate its cluster center:

[0088]

[0089] Where t represents the t-th iteration; |c i | indicates that it belongs to cluster c i The number of reference points;

[0090] (5) Define the loss function

[0091]

[0092] Repeat steps (3) and (4) until all cluster centers are fixed, i.e., the loss function J < ε, where ε is the loss threshold.

[0093] (6) Rewrite the converged cluster centers in matrix form i dim×N , where dim is the output dimension of the autoencoder;

[0094] Step 4: Define the similarity criteria between location fingerprints

[0095] The compressed location fingerprint retains the main user location information, and the Euclidean distance between location fingerprints is defined as the fingerprint similarity criterion:

[0096]

[0097] Among them, the closer the Euclidean distance between location fingerprints, the higher the similarity between the fingerprints;

[0098] Step 5: Extract user fingerprint information

[0099] The user channel state information is estimated using the pilot signal transmitted by user k, denoted as . The angle domain channel is obtained by performing a Fourier transform. Then The compressed data is extracted from the hidden layer of the trained autoencoder and used as the user's location fingerprint, denoted as Θ. k ;

[0100] Step 6: Fingerprint matching and location estimation

[0101] First, calculate the user location fingerprint Θ according to the similarity judgment criteria obtained in step 4. k The similarity Λ(Θ) between each fingerprint cluster center k ,μ iSelect the two clusters with the highest similarity, denoted as Ξ; then compare the user's fingerprint with the reference point fingerprints within Ξ one by one, and select the three nearest neighbors with the highest similarity, denoted as Ξ. The corresponding similarity coefficient is denoted as

[0102] Step 7: Estimate user location

[0103] The Weighted K-Nearest Neighbors (JNNN) algorithm, an improvement on the KNNN algorithm, achieves more accurate localization using the same number of nearest neighbors (RPs). Its formula is expressed as:

[0104]

[0105] In the above embodiment, the service area is considered to be 100×100m. 2 The decellularized massive MIMO system has a sampling interval of 5 meters between adjacent reference points and 441 reference points. There are 20 users to be located, 10 access points (APs), 32 antenna array elements, 10 scattering paths, a single-sided angular spread of 6°, an uplink pilot signal transmission power of 100mW, 100 channel implementations, 100 random implementations of AP and user locations, 20 clusters, and 3 selected reference points. The autoencoder has an input and output layer with 32 neurons each, and three hidden layers with 8, 3, and 8 neurons respectively. The angular domain channel dimension is 32*10, and the compressed fingerprint dimension is 3*10.

[0106] In this embodiment, the implementation process of the proposed technical solution is as follows: Figure 1 As shown, each operation corresponds one-to-one with steps 1-7 in the autoencoder-assisted decellularized large-scale MIMO low-dimensional fingerprint localization method, and will not be repeated here. The performance simulation of user localization using this technical solution is as follows: Figure 4 , Figure 5 As shown. Figure 4 This study reveals the impact of this technical solution on positioning performance compared to other data dimensionality reduction algorithms. Figure 5 The localization performance of this technical solution was compared with that of other clustering algorithms, which verified that the proposed technical solution has the best performance.

[0107] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. An autoencoder-assisted decellularized large-scale MIMO low-dimensional fingerprint localization method, characterized in that: Includes the following steps: Step 1: Extract the angle information matrix of the reference point user Assuming a cellular-free massive MIMO system Access points (APs) are randomly distributed within the service area. Each AP is equipped with a uniform linear antenna array, denoted by M elements. The service area is uniformly divided, and the intersections of the grids serve as reference points, denoted by R. All APs estimate the channel using uplink pilot signals to obtain the CSI at each reference point. Considering a narrowband multipath channel model, the... The AP and the first The channel between RPs is: （1） in: To determine the number of resolvable scattering paths, we take L=10; For the first Complex gain of the path, The large-scale fading coefficient, comprising path loss and shadow fading, is expressed as: （2） in: For the first The Euclidean distance between the r-th AP and the r-th RP Indicates shadow decay; among which, The variance of shadow fading is taken as the standard value; during simulation verification, a uniform value is used. ; Indicates the first The array response vector corresponding to each scattering path. ,in Antenna spacing, For carrier wavelength, For the first The signal angle of arrival for each path; No. The uplink pilot signal received at each AP from the r-th RP is: （3） in: Indicates the uplink pilot sequence. This represents an additive white Gaussian noise vector; The channel can be estimated based on the received uplink pilot signal, denoted as: ; Performing a Fourier transform on the estimated channel yields the angular domain channel: （4） in: Represents the DFT matrix, ; The angle-domain channel matrix is ​​represented as: （5） To suppress fingerprint fluctuations caused by small-scale fading, the angle-domain channel matrix across multiple coherent time intervals is further processed as follows: （6） Step 2: Train the autoencoder network The result obtained in step 1 As a training dataset, it is input into the autoencoder network to learn the intrinsic relationship between the channel information matrix and the user's location, and to extract low-dimensional fingerprint information; input data The data is sent to the input layer, passed through one or more hidden layers, and then to the output layer to obtain the output data. Automatic encoders reduce the amount of input data. With output data The error between the two is used to learn the inherent features hidden in the original data, i.e., the output matrix of the hidden layer. After the network training is complete, the hidden layer output will be in matrix form. Fingerprint as a reference point location; Step 3: Fingerprint clustering based on K-means++ (1) Randomly select Q reference point locations and their vectorized fingerprints as cluster centers, denoted as . ; (2) Define Euclidean distance ; 3) For each reference point location fingerprint Calculate its relationship with each cluster center. The Euclidean distance is used to assign the fingerprint of the reference point to the cluster corresponding to the minimum value of the Euclidean distance; (4) For each cluster, recalculate its cluster center: （7） Where t represents the t-th iteration; Indicates belonging to a cluster The number of reference points; (5) Define the loss function (8) Repeat steps (3) and (4) until all cluster centers are fixed, i.e., the loss function is obtained. ,in This is the loss threshold. (6) Rewrite the converged cluster centers in matrix form. , where dim is the output dimension of the autoencoder; Step 4: Define the similarity criteria between location fingerprints The compressed location fingerprint retains the main user location information, and the Euclidean distance between location fingerprints is defined as the fingerprint similarity criterion: （7） Among them, the closer the Euclidean distance between location fingerprints, the higher the similarity between the fingerprints; Representative reference point r The fingerprint information comes from the first n Information about each access point Representative reference point The fingerprint information comes from the first n Information about each access point; Step 5: Extract user fingerprint information The user channel state information is estimated using the pilot signal transmitted by user k, denoted as . Perform a Fourier transform to obtain the angle domain channel. Then The compressed data is extracted from the hidden layer of the trained autoencoder and used as the user's location fingerprint, denoted as [fingerprint]. ; Step 6: Fingerprint matching and location estimation First, calculate the user's location fingerprint according to the similarity judgment criteria obtained in step 4. Similarity with the center of each fingerprint cluster Select the two clusters with the highest similarity and denote them as: Then the user's fingerprint and The fingerprints of each reference point within the range are compared one by one, and the three nearest neighbors with the highest similarity are selected and denoted as . The corresponding similarity coefficient is denoted as ; Step 7: Estimate user location The Weighted K-Nearest Neighbors (JNNN) algorithm, an improvement on the KNNN algorithm, achieves more accurate localization using the same number of nearest neighbors (RPs). Its formula is expressed as: （8）。

Images