Global positioning method based on deep learning difference matrix measurement to construct channel graph

Abstract

The application belongs to the technical field of mobile communication, and discloses a global positioning method for constructing a channel graph based on a difference matrix measurement of deep learning, wherein channel state information data is collected and processed, a difference measurement matrix based on deep learning is constructed, a channel graph is constructed by using multi-dimensional scaling manifold learning technology, and global positioning is realized by combining rotation, conformal transformation and affine transformation, so that the universality of the method for constructing a channel graph and global positioning based on a difference measurement model prediction value is ensured.

Description

Technical Field

[0001] This invention belongs to the field of mobile communication technology, specifically relating to a global localization method for constructing a channel graph based on deep learning difference matrix metric. Background Technology

[0002] To cope with the significant increase in traffic volume, number of terminals, and number of users, and the demand for high-reliability and low-latency communication, future wireless communication systems are designed with massively multi-input multiple-output (mMIMO) technology. mMIMO technology significantly enhances wireless signal coverage by multiplying system communication capacity and utilizing techniques such as spatiotemporal beamforming, multi-user separation, and data precoding. Base stations (BSs) utilize mMIMO antenna technology to collect rich channel state information (CSI) from user equipment (UEs). By leveraging the similarity of this information in a high-dimensional space and employing a self-supervised learning strategy, a two-dimensional or even three-dimensional map of the radio environment (hereinafter referred to as: channel map) can be drawn. Global positioning of user equipment (UEs) can be achieved based on the channel map without relying on location tags or with very few location tags.

[0003] In recent years, scholars both domestically and internationally have conducted in-depth research on channel map localization techniques. Typical methods include: the localization method proposed by Jaakko Pihlajasalo et al. in 2020 based on multi-base station channel map generation, combination, and mapping transformation; and the localization method proposed by Florian Euchne et al. in 2023 that fuses channel state information with classical source localization (triangulation and polygon localization) information. In these efficient channel map construction and localization methods, the measurement of channel state information dissimilarity is a key technology. Scholars both domestically and internationally have also proposed a series of methods, including: the timestamp-based dissimilarity measurement method proposed by P. Ferrand et al. in 2021; the cosine similarity-based measurement method proposed by L. LeMagoarou in 2021; the dissimilarity measurement method based on channel impulse response amplitude (CIRA) proposed by M. Stahlke et al. in 2023; and the dissimilarity matrix calculation method based on angular delay curves (ADP) proposed by Phillip Stephan et al. in 2024.

[0004] However, the difference measurement of the aforementioned channel state information depends on specific channel characteristics such as amplitude attenuation and phase change. This makes the channel map-based user equipment positioning method usually have the following drawbacks: the positioning method based on specific characteristics such as amplitude attenuation and phase change is highly dependent on the environment, making it difficult to ensure the universality of the positioning technology. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a global positioning method for constructing a channel graph based on a deep learning difference matrix metric. By using a deep learning-based difference metric matrix, a channel graph is constructed using multidimensional scaling manifold learning techniques, and combined with transformation methods to achieve global positioning, ultimately achieving universally applicable and accurate positioning of user equipment (UE) targets.

[0006] The global localization method for constructing a channel map based on deep learning difference matrix metric, as described in this invention, includes the following steps:

[0007] Step 1: Collect channel state data and perform time-domain transformation on the collected channel state data to obtain time-domain channel state information data;

[0008] Step 2: Perform feature engineering on the time-domain channel state information data, and construct a multi-base station difference matrix based on deep learning differences;

[0009] Step 3: Based on the multi-base station difference matrix, construct the channel graph using multidimensional scaling manifold learning;

[0010] Step 4: Perform rotation transformation, commonality transformation and affine transformation on the channel map to achieve global positioning.

[0011] Furthermore, step 1 specifically includes:

[0012] Step 1-1: Obtain channel state information data;

[0013] Given a base station BS consisting of B distributed antenna arrays mMIMO, each array equipped with M antennas, and a user equipment UE with a single antenna, the base station BS contains B×M antennas, and the distance between the base station BS and the single antenna of the user equipment UE is measured at the l-th discrete time as N. sub The channel coefficients of each Orthogonal Frequency Division Multiplexing (OFDM) subcarrier are expressed as a frequency domain tensor H. (l) The formula is:

[0014]

[0015] Among them, the set of L discrete-time channel coefficients and their corresponding measurement times constitute the frequency domain channel state information dataset (CSI). Fre The formula is:

[0016] CSI Fre ={(H (l) ,t (l) )},l=1,...,L (2)

[0017] Among them, t (l) ∈R represents the discrete timestamp of the channel coefficient measurement, in milliseconds; R is the set of real numbers;

[0018] Steps 1-2: Time-domain transformation of the channel state information dataset:

[0019] For channel map-based positioning applications, a time-domain transformation of the channel state dataset is required, transforming H... (l) Performing a unit discrete Fourier inverse transform along the frequency axis yields the corresponding time-domain tensor. The formula is:

[0020]

[0021] in, n is the order of τ, im is the imaginary unit, τ is the time tap, τ = 1 corresponds to the first time tap, and τ = N sub Corresponding to the last time tap;

[0022] Obtain the corresponding time-domain channel state information (CSI) data. Tim The formula is:

[0023]

[0024] Furthermore, in step 2, feature engineering is performed on the time-domain channel state information data, specifically as follows:

[0025] Step 2-1-1: Calculate the autocorrelation characteristics of the time-domain channel state information data;

[0026] Given a base station BS consisting of B distributed antenna arrays mMIMO, each array equipped with M antennas, and a minimum time tap τ, min and maximum time tap τ max Based on the Cartesian product, an index tuple J is obtained, as shown in the formula:

[0027] J = {1, ..., B} 2 ×{1,...,M} 2 ×{τ min ,...,τ max}(5)

[0028] Where, 1 << τ min <N sub / 2<τ max <<N sub That is, only the time taps of multiple channel impulse responses (CIRs) including the visible path and the first few path components are considered; τ min and τ max The choice depends on the expected maximum delay distribution;

[0029] For each pair of antenna arrays b in the base station BS with antennas i ,b jEach pair of antennas m p ,m q And the time-domain tensor of each time tap τ, l discrete-time measurements. The autocorrelation feature c between samples (l) The calculation formula is:

[0030]

[0031] in, (b i ,b j ,m p ,m q ,τ)∈J;i,j,p,q∈R,R is the set of real numbers;* denotes a pair of τ)∈J;i,j,p,q∈R,R is the set of real numbers;* represents a pair of τ)∈J;i,j,p,q∈R;i,j,p,q∈R;R is the set of real numbers;* represents a pair of τ)∈J;i,j,p,q∈R;R is the set of real numbers;* represents a Take conjugate;

[0032] Step 2-1-2: Obtain the eigenvectors by autocorrelation eigenvectorization;

[0033] By c (l) The real and imaginary parts of the vector are separated and vectorized to obtain the corresponding eigenvector f. (l) The formula is:

[0034]

[0035] Where Re{} and Im{} represent complex numbers c, respectively. (l) The real and imaginary parts.

[0036] Furthermore, in step 2, a multi-base station difference matrix is ​​constructed based on the differences learned through deep learning, specifically as follows:

[0037] Step 2-2-1: Train and optimize the differential measurement model based on deep neural network (DNN);

[0038] eigenvectors f of any two discrete time i and j (i) f (j) As the model input, where 1≤i,j≤L, and i≠j; through processing by multiple hidden layers, the output layer obtains the estimated value of the difference between channel state information data acquisition times i and j. Calculate the output estimate Compared with the actual measured value d time,i,j =|t (i) -t (j) Error between |

[0039] The loss function is defined using the normalized mean square error (NMSE), and the formula is:

[0040]

[0041] Where, dtime,i,j =|t (i) -t (j) | represents the actual difference between channel state information data acquisition times i and j, and β > 0 is a hyperparameter;

[0042] Finally, the gradient is calculated using the loss function, and the optimized difference metric model DΘ is obtained based on backpropagation and stochastic gradient descent.

[0043] Step 2-2-2: Based on training, optimize the difference measurement model DΘ to predict and output the difference values ​​of all channel state information data acquisition times i and j. as well as Based on the symmetry of dissimilarity, the results are averaged to obtain the difference d between the channel state information data acquired at acquisition times i and j. DL,i,j The formula is:

[0044]

[0045] Step 2-2-3: Construct a multi-base station difference matrix based on visibility;

[0046] Based on the differences in the data acquisition times i and j of all channel state information, the difference measurement matrix D for the base station in L time periods is obtained. DL The formula is:

[0047]

[0048] Given n base stations (BS) and one user equipment (UE), if the UE is visible to base station k for both discrete measurement times i and j (1 ≤ k ≤ n), then the visibility of the UE to base station k is defined as follows: otherwise

[0049] The difference measurement matrix D between n base stations (BS) and user equipment (UE) is based on a weighted average combination. DL The elements in the matrix yield the global dissimilarity metric matrix D. DLG The formula is:

[0050]

[0051] in, It is the difference between the channel state information data acquisition times i and j based on the output of the trained model; It is the difference value between the channel state information data acquisition time j and i based on the output of the trained model; It is the difference between the channel state information data of base station k and user equipment UE at two discrete measurement times i and j; It refers to the visibility of the user equipment (UE) to the base station k at two discrete measurement times i and j; d DLG,i,jIt is the time-domain tensor of the channel state information data collected at times i and j. In higher dimensions Mid-range.

[0052] Furthermore, step 3 specifically involves:

[0053] Given the time-domain tensor of time-domain channel state data at any two discrete acquisition times i and j The multidimensional scaling manifold learning method is used to obtain the two in the reduced-dimensional space R. D＝2 Mapping position z (i) , z (j) , that is:

[0054]

[0055] The mapping calculation satisfies the following conditions:

[0056] ||z (i) -z (j) ||2∝d DLG,i,j (13)

[0057] Among them, ||z (i) -z (j) ||2 is the time-domain tensor In the corresponding dimensionality reduction space R D＝2 Mapping position z (i) ,z (j) The distance between them, ∝ indicates the closest possible proximity;

[0058] For all time-domain channel state data of discrete time intervals l = 1, ..., L, D DLG Midpoint to distance d DLG,i,j Corresponding to ||z (i) -z (j) The distance between ||2 should satisfy the condition of minimizing the mean square distance, that is:

[0059]

[0060] All z that satisfy the conditions (l) Composed in the reduced-dimensional space R D＝2 Channel state data channel diagram Z cc , that is:

[0061]

[0062] Furthermore, the channel state data channel diagram in the reduced-dimensional space is constructed, and the specific steps are as follows:

[0063] (1) Using D DLG Element data calculation

[0064] (2) Calculate the inner product And further, the dimensionality-reduced inner product matrix is ​​obtained. in, It is D DLG The mean of all elements in the i-th row of the expression. It is D DLG The mean of all elements in the j-th row of the expression. It is D DLG The global mean;

[0065] (3) Perform eigenvalue decomposition on matrix IPM: IPM = VΛV T Where T represents the transpose of the eigenvector matrix V, It is a diagonal matrix composed of eigenvalues. When V has two non-zero eigenvalues, V = V * , Λ=Λ * =diag(λ1,λ2), then the time-domain channel state data channel diagram is obtained.

[0066] Furthermore, step 4 specifically involves:

[0067] Step 4-1: When the base station collects channel state data with the user equipment at L discrete time intervals, it synchronously acquires the reference signal received power dataset and estimates the position of the base station in the channel map.

[0068] Step 4-2: Given the map spatial position of each base station center translation, calculate its rotation angle relative to the estimated position in the channel map, and use the circular average method to estimate the global rotation angle.

[0069] Step 4-3: Perform conformal transformation using elliptical mesh mapping to map circles to squares to eliminate roundness in the channel graph;

[0070] Step 4-4: Construct the homogeneous coordinate matrix, obtain the affine transformation matrix based on the linear equation system, and solve for it;

[0071] Steps 4-5: Perform global positioning based on affine matrices.

[0072] The beneficial effects of this invention are as follows: The method described in this invention accurately captures specific signal features in complex environments by constructing a difference metric matrix based on deep learning. It then utilizes multidimensional scaling manifold learning techniques to construct a channel map and combines rotation, conformal, and affine transformations to achieve global localization. This invention designs a deep neural network (DNN) difference metric model that uses discrete-time time-domain channel state information as model input and optimizes it through backpropagation and stochastic gradient descent. This model uses the time difference of real measurements, independent of specific scene features, as a verification value, ensuring the universality of the method for constructing channel maps and performing global localization based on the predicted values ​​of the difference metric model. Attached Figure Description

[0073] Figure 1 This is a flowchart of the method described in this invention.

[0074] Figure 2 This is a schematic diagram of the channel coefficients between the base station (BS) and the user equipment (UE);

[0075] Figure 3 This is a flowchart of the training process for a differential measurement model based on deep neural networks;

[0076] Figure 4 This is a schematic diagram of the optimization model structure based on backpropagation calculation and stochastic gradient descent;

[0077] Figure 5 This is a diagram illustrating the visibility of a user equipment (UE) to a single base station (BS).

[0078] Figure 6 This is a diagram illustrating the visibility of the user equipment (UE) to base stations (BS1) and (BS2);

[0079] Figure 7 It is a channel graph constructed based on multidimensional scaling manifold learning;

[0080] Figure 8 This is a schematic diagram showing the estimated spatial locations of the four base stations in the channel map;

[0081] Figure 9 This is a schematic diagram of the global location of conformal transformation points in the map space based on the affine transformation matrix in the channel graph. Detailed Implementation

[0082] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0083] like Figure 1 As shown, the global localization method for constructing a channel map based on deep learning difference matrix metric according to the present invention includes the following steps:

[0084] Step 1: Collect channel state data and perform time-domain transformation on the collected channel state data to obtain time-domain channel state information data;

[0085] Step 2: Perform feature engineering on the time-domain channel state information data, and construct a multi-base station difference matrix based on deep learning differences;

[0086] Step 3: Based on the multi-base station difference matrix, construct the channel graph using multidimensional scaling manifold learning;

[0087] Step 4: Perform rotation transformation, commonality transformation and affine transformation on the channel map to achieve global positioning.

[0088] Step 1 specifically includes:

[0089] Step 1-1: Obtain channel state information data;

[0090] Given a base station (BS) consisting of B distributed antenna arrays (mMIMO), each array equipped with M antennas, and a user equipment (UE) with a single antenna, the base station (BS) contains B×M antennas, and the N values ​​measured between the base station (BS) and the single antenna of the user equipment (UE) at the l-th discrete time are... sub The channel coefficients of each Orthogonal Frequency Division Multiplexing (OFDM) subcarrier are expressed as a frequency domain tensor H. (l) ,like Figure 2 As shown, the formula is:

[0091]

[0092] Among them, the set of L discrete-time channel coefficients and their corresponding measurement times constitute the frequency domain channel state information dataset (CSI). Fre The formula is:

[0093] CSI Fre ={(H (l) ,t (l) )},l=1,...,L (2)

[0094] Among them, t (l) ∈R represents the discrete timestamp of the channel coefficient measurement, in milliseconds; R is the set of real numbers;

[0095] Steps 1-2: Time-domain transformation of the channel state information dataset:

[0096] H (l) Performing a unit discrete Fourier inverse transform along the frequency axis yields the corresponding time-domain tensor. The formula is:

[0097]

[0098] in, n is the order of τ, im is the imaginary unit, τ is the time tap, τ = 1 corresponds to the first time tap, and τ = N sub Corresponding to the last time tap;

[0099] Obtain the corresponding time-domain channel state information (CSI) data. Tim The formula is:

[0100]

[0101] In step 2, a differential measurement model based on a deep neural network is trained. The basic process is as follows: Figure 3 As shown.

[0102] (1) Characterization of channel state information data;

[0103] 1) Calculate the autocorrelation characteristics of time-domain channel state information data

[0104] Given a base station (BS) consisting of B distributed antenna arrays (mMIMO), each array equipped with M antennas, and a minimum time tap τ, min and maximum time tap τ max Based on the Cartesian product, an index tuple J is obtained, as shown in the formula:

[0105] J = {1, ..., B} 2 ×{1,...,M} 2 ×{τ min ,...,τ max}(5)

[0106] Where, 1 << τ min <N sub / 2<τ max <<N sub That is, only the time taps of multiple channel impulse responses (CIRs) including the visible path and the first few path components are considered. τ min and τ max The choice depends on the expected maximum delay distribution.

[0107] For each pair of antenna arrays b in the base station (BS) of the antenna i ,b j Each pair of antennas m p ,m q And the time-domain tensor of each time tap τ, l discrete-time measurements. The autocorrelation feature between samples is c (l) The formula is:

[0108]

[0109] in (b i ,b j ,m p ,m q ,τ)∈J;i,j,p,q∈R,R is the set of real numbers;* denotes a pair of τ)∈J;i,j,p,q∈R,R is the set of real numbers;* represents a pair of τ)∈J;i,j,p,q∈R;i,j,p,q∈R;R is the set of real numbers;* represents a pair of τ)∈J;i,j,p,q∈R;R is the set of real numbers;* represents a Take conjugate;

[0110] 2) Autocorrelation eigenvectorization yields eigenvectors;

[0111] By c (l) The real and imaginary parts of the vector are separated and vectorized to obtain the corresponding eigenvector f. (l) The formula is:

[0112]

[0113] Re{} and Im{} represent the complex number c, respectively. (l) The real and imaginary parts.

[0114] (2) Training of a differential measurement model based on deep neural networks (DNN)

[0115] 1) Model training optimization based on stochastic gradient descent and backpropagation calculations;

[0116] The eigenvectors f of any two discrete time i and j (1≤i, j≤n and i≠j) (i) and f (j) As input to the model, the difference between channel state information data acquisition times i and j is obtained from the output layer through processing by multiple hidden layers. Calculate the output estimate Compared with the actual measured value d time,i,j =|t (i) -t (j) Error between |

[0117] The loss function is defined using the normalized mean square error (NMSE), and the formula is:

[0118]

[0119] Where, d time,i,j =|t (i) -t (j) | represents the actual difference between channel state information data acquisition times i and j, and β > 0 is a hyperparameter;

[0120] Finally, the gradient is calculated using the loss function, and the optimized model DΘ is obtained based on backpropagation and stochastic gradient descent. The basic network structure is as follows: Figure 4As shown.

[0121] In this example, the differential measurement model DΘ consists of one input layer, three hidden layers, and one output layer; the number of neurons in the three hidden layers are 128, 64, and 32, respectively, all using the ReLU activation function, and the nodes in each layer are fully connected.

[0122] 2) The DΘ-based training-based optimization model predicts the difference between the data acquisition times i and j for all channel state information. as well as Based on the symmetry of dissimilarity, the results are averaged to obtain the difference d between the channel state information data acquired at acquisition times i and j. DL,i,j The formula is:

[0123]

[0124] (3) Construction of a multi-base station difference matrix based on visibility;

[0125] 1) Based on the differences in the data acquisition times i and j of all channel state information, the difference measurement matrix D for the base station in L time periods is obtained. DL The formula is:

[0126]

[0127] Given n base stations and one user equipment (UE), if the UE is visible (LoS) to base station k for both discrete measurement times i and j, 1 ≤ k ≤ n, then the visibility of the UE to base station k is defined as follows: otherwise

[0128] In this example, Figure 5 This is an example of user equipment (UE) visibility to a single base station (BS), where Figure 5 In (a) This indicates that the user equipment (UE) is visible to the base station (BS) at both discrete measurement times i and j. Figure 5 (b) This indicates that the user equipment (UE) is visible to the base station (BS) at discrete measurement time i, but not visible to the base station (BS) at discrete measurement time j.

[0129] Figure 6 This is an example of user equipment (UE) visibility to multiple base stations (BS), where Figure 6 In (a) This indicates that the user equipment (UE) is visible to the base station (BS1) at discrete measurement time i, but not visible to the base station (BS1) at discrete measurement time j. This indicates that the user equipment (UE) is visible to the base station (BS2) at discrete measurement times i and j. Figure 6 (b) This indicates that the user equipment (UE) is visible to the base station (BS1) at discrete measurement time i, but not visible to the base station (BS1) at discrete measurement time j. This indicates that the user equipment (UE) is not visible to the base station (BS2) at discrete measurement time i, but is visible to the base station (BS2) at discrete measurement time j.

[0130] 2) A weighted average combination of the difference metric matrix D between user equipment (UE) and n base stations (BS). DL The elements in the matrix yield the global dissimilarity metric matrix D. DLG The formula is:

[0131]

[0132] in, It is the difference between the channel state information data acquisition times i and j based on the output of the trained model; It is the difference value between the channel state information data acquisition time j and i based on the output of the trained model; It is the difference between the channel state information data of base station k and user equipment UE at two discrete measurement times i and j; It refers to the visibility of the user equipment (UE) to the base station k at two discrete measurement times i and j; d DLG,i,j It is the time-domain tensor of the channel state information data collected at times i and j. In higher dimensions Mid-range.

[0133] In step 3, given the time-domain tensor of the time-domain channel state data at any two discrete acquisition times i and j The multidimensional scaling manifold learning method is used to obtain the two in the reduced-dimensional space R. D ＝2 Mapping position z (i) , z (j) , that is:

[0134]

[0135] The mapping calculation satisfies the following conditions:

[0136] ||z (i) -z (j) ||2∝d DLG,i,j (13)

[0137] Among them, ||z (i) -z (j) ||2 is the time-domain tensor In the corresponding dimensionality reduction space RD＝2 Mapping position z (i) ,z (j) The distance between them, ∝ indicates the closest possible proximity.

[0138] For all time-domain channel state data of discrete time intervals l = 1, ..., L, D DLG Midpoint to distance d DLG,i,j Corresponding to ||z (i) -z (j) The distance between ||2 should satisfy the condition of minimizing the mean square distance, that is:

[0139]

[0140] All z that satisfy the conditions (l) Composed in the reduced-dimensional space R D＝2 Channel state data channel diagram Z cc , that is:

[0141]

[0142] The specific implementation steps are as follows:

[0143] (1) Using D DLG Element data calculation

[0144] (2) Calculate the inner product And further, the dimensionality-reduced inner product matrix is ​​obtained. in, It is D DLG The mean of all elements in the i-th row of the expression. It is D DLG The mean of all elements in the j-th row of the expression. It is D DLG The global mean.

[0145] (3) Perform eigenvalue decomposition on matrix IPM: IPM = VΛV T Where T represents the transpose of the eigenvector matrix V, It is a diagonal matrix composed of eigenvalues. When V has two non-zero eigenvalues, V = V * , Λ=Λ * =diag(λ1,λ2), then the time-domain channel state data channel diagram is obtained.

[0146] In this example, Figure 7 The channel map is obtained through steps 1 to 3, based on 10,000 measurement time (in milliseconds) of channel state data measured by 4 base stations and 1 user equipment (UE).

[0147] In step 4, the channel diagram Z obtained based on the channel state data... cc Lacking true scale and shape, it cannot be used for map spatial positioning and requires rotational transformation, conformal transformation, and affine transformation. The specific steps include:

[0148] (1) Location estimation of base station (BS) in the channel graph;

[0149] 1) When the base station (BS) collects channel state data with the user equipment (UE) at L discrete time intervals, it synchronously acquires the reference signal received power dataset RSRQs, which is defined as:

[0150] RSRQs={power1,power2,...,power L}(16)

[0151] Among them, power i ,1≤i≤L, represents the received signal power at the i-th discrete time.

[0152] 2) Sort the elements in CSI-RSRQs in descending order to obtain a subset RSRQs consisting of the first T elements. t Its definition is:

[0153] RSRQs T ={por1,por2,...,por T}(17)

[0154] 3) Channel diagram of time-domain channel state data generated from base station (BS) and user equipment (UE) based on the acquisition time of elements in RSRQst. Get a set of T corresponding mapping positions Its definition is:

[0155]

[0156] 4) Based on The mean of the elements in the middle is used to estimate the base station (BS) in the channel graph Z. cc Position z in (bs) Its definition is:

[0157]

[0158] In this example, Figure 8 This is a schematic diagram showing the estimated spatial locations of the four base stations in the channel map.

[0159] (2) Rotation transformation;

[0160] 1) Center shift of the base station's (BS) map spatial location;

[0161] Calculate the average center position of n base stations (BS), and subtract the coordinates of the average center position from the coordinates of each base station (BS) to achieve a center shift in the spatial location of the base station (BS) on the map. The calculation formula is as follows:

[0162]

[0163] in, It is the spatial location of the base station k on the map, BSS cen It is the average center position of the spatial locations of n base stations (BS) on the map. It is the position shifted to the average center.

[0164] 2) Calculation of the rotation angle between the map position of the base station BS center translation and the estimated position of the base station BS in the channel map.

[0165] Given the map spatial location of base station k after center translation And the estimated location of base station k in the channel graph The rotation angle θ between two points is calculated using the two-parameter arctangent function, and the formula is as follows:

[0166]

[0167] in, These are the estimated locations of base station (BS) k in the channel graph. x and y coordinates These are the map spatial locations after the center of base station (BS) k is shifted. The horizontal and vertical coordinates.

[0168] Furthermore, based on the location information of n base stations and the calculated n rotation angles, the global rotation angle is estimated using the circular average method. The formula is:

[0169]

[0170] Where, θ j The rotation angle is calculated based on the location information of base station j.

[0171] 3) Use global estimation of rotation angle right Each mapping position z in (l) Rotate to obtain the corresponding rotation point. The formula is:

[0172]

[0173] Furthermore, we obtain the set of all rotation points.

[0174] (3) Conformal transformation;

[0175] Conformal transformation is performed using elliptical mesh mapping to map circles to squares in order to eliminate roundness in the channel graph;

[0176] 1) For middle Scaling is performed based on the maximum distance from the point to the origin of the channel map center, calculated using the following formula:

[0177]

[0178] in, Indicates the point of rotation Distance to the origin of the channel map center This represents the maximum distance from the L rotation points to the origin. This represents the scaled-down points;

[0179] 2) For the scaled points Conformal points are obtained by conformal mapping of elliptical meshes. The calculation formula is:

[0180]

[0181] in, These are the scaling points. x and y coordinates Corresponding to conformal back points The horizontal and vertical coordinates.

[0182] Furthermore, we obtain the set of all conformal points.

[0183] (4) Affine transformation;

[0184] Given a set of spatial locations of n base stations (BS) And its corresponding position in the channel diagram Position set after rotation, scaling, and conformal processing The affine transformation parameters are solved as follows:

[0185] 1) Construct the homogeneous coordinate matrix;

[0186] For each point and the corresponding target point BS k • Maplocs, 1≤k≤n, is expanded to homogeneous coordinates and arranged into the following matrix by rows:

[0187]

[0188] 2) Construct a system of linear equations;

[0189] For each point and the corresponding target point BS k • Maplocs, 1≤k≤n, satisfy the following relationship:

[0190]

[0191] For all n pairs of points, the relationship can be expressed as a matrix equation:

[0192]

[0193] Where M is the affine transformation matrix.

[0194] 3) Solve for the affine transformation matrix;

[0195] Solve for the affine transformation matrix M based on the following matrix equation:

[0196] M = (X T X) -1 X T Y(29)

[0197] Among them, X T This is the transpose of the matrix.

[0198] (5) Global positioning based on affine matrices;

[0199] Given the set of conformal transformation points in the channel graph And the affine transformation matrix M, for any conformal transformation point Construct homogeneous coordinates:

[0200]

[0201] Furthermore, based on homogeneous coordinates The global location point Maploc in the corresponding map space is calculated using the affine transformation matrix M. l The calculation formula is:

[0202]

[0203] Finally, the set of global localization points of all conformal transformation points in the channel graph in the map space is obtained.

[0204]

[0205] In this example, Figure 9 This is a schematic diagram of the transformation from a channel map with 30*75 fixed points to map space, based on the affine transformation matrix of the conformal transformation matrix of the channel map.

[0206] Error analysis was performed on the coordinates of 30*75 points based on the affine transformation of the channel map and the true coordinates. The results are shown in Table 1. Analysis of the experimental results of the sampling point residuals and total errors in Table 1 shows that the affine transformation model of the channel map obtained from 10 sample points exhibits high positioning accuracy when applied to 30*75 test points, verifying the universality of the global positioning method for constructing the channel map based on deep learning difference matrix metric proposed in this invention.

[0207] Table 1 Error Analysis of Map Space Based on Affine Transformation of Channel Graph

[0208]

[0209] The above description is merely a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made based on the description and drawings of the present invention are within the protection scope of the present invention.

Claims

1. A global positioning method based on deep learning difference matrix measurement to construct a channel graph, characterized in that, The method comprises the following steps: Step 1, collecting channel state data, and performing time domain transformation on the collected channel state data to obtain time domain channel state information data; Step 2, performing feature engineering on the time domain channel state information data, and constructing a multi-base station difference matrix based on the difference of deep learning; wherein the difference of deep learning refers to using a difference measurement model based on a deep neural network DNN optimized by training to predict the difference value of all channel state information data collection times i, j, and based on the symmetry of the dissimilarity, the result is averaged to obtain the difference measurement of the channel state information data of the collection times i, j; Step 3, based on the multi-base station difference matrix, a channel graph is constructed by using multi-dimensional scaling manifold learning; Step 4, performing rotation transformation, common transformation and affine transformation on the channel graph to realize global positioning. 2.The global positioning method of constructing channel graph based on deep learning difference matrix measurement according to claim 1, wherein, Step 1 specifically comprises: Step 1-1, obtaining channel state information data; Given a base station BS composed of distributed antenna arrays mMIMO and equipped with antennas per antenna array and a user equipment UE with a single antenna, the base station BS comprises antennas whose channel coefficients between the single antenna of the user equipment UE at the discrete time measurements are expressed as a frequency domain tensor of orthogonal frequency division multiplexing subcarriers, the formula is: (1) wherein a set of channel coefficients of the discrete-time measurements and the corresponding measurement times form a frequency domain channel state information data set , which is (2) wherein, is a discrete time stamp for the channel coefficient measurement in milliseconds; is a real number set; Step 1-2, time domain transformation of the channel state information data set: The following The inverse unitary discrete Fourier transform along the frequency axis gives the corresponding time-domain tensor The formula is: (3) wherein , is in the order of is the imaginary unit, is the time tap, corresponds to the first time tap, corresponds to the last time tap; corresponding time domain channel state information data is obtained with the formula: (4)。 3.The global positioning method of constructing channel graph based on deep learning difference matrix metric according to claim 2, characterized in that, In step 2, the feature engineering of the time domain channel state information data specifically comprises: Step 2-1-1, calculating the autocorrelation feature of the time domain channel state information data; Given a base station BS composed of distributed antenna arrays mMIMO and each antenna array equipped with antennas, and a minimum time tap and a maximum time tap , a set of index tuples is computed based on Cartesian product, formula is: (5) wherein i.e. only time taps of the channel impulse response CIR of the multiple channels comprising the visual path and the first multiple path components are considered; and The selection of the number of taps depends on the expected maximum delay profile; For each pair of antenna arrays in a base station BS , each pair of antennas and each time tap , sample autocorrelation features between the time-domain tensors of discrete-time measurements are computed as (6) wherein ; is the set of real numbers; * denotes taking the complex conjugate; the conjugate; Step 2-1-2, vectorizing the autocorrelation feature to obtain a feature vector; Vectorization is achieved by separating the real and imaginary parts of , resulting in the corresponding eigenvectors , where the formula is: (7) in, , Representing complex numbers respectively The real and imaginary parts. 4.The global positioning method of constructing channel graph based on deep learning difference matrix measurement according to claim 3, characterized in that, In step 2, the multi-base station difference matrix is constructed based on the difference of deep learning, specifically: Step 2-2-1, training and optimizing the difference measurement model based on the deep neural network DNN; Eigen vectors of any two discrete times i, j , as model input, where 1≤i,j≤L and i≠j; by processing through multiple hidden layers, an estimate of the difference in channel state information data collection times i, j is obtained from the output layer ; the error between the output estimate and the true measured value is calculated ; A loss function is defined by normalizing the mean square error MSE, and the formula is: (8) wherein, denotes the true difference of the channel state information data collection times i, j, is a hyperparameter; Finally, the gradient is calculated by the loss function, and the optimized difference measure model is obtained based on back propagation calculation and stochastic gradient descent ; Step 2-2-2, training-based optimization of the dissimilarity metric model Predicted output all-channel state information data collection time i, j dissimilarity value And And based on the symmetry of dissimilarity, the result is averaged to obtain the dissimilarity metric of the channel state information data of the collection time i, j , the formula is: (9) Step 2-2-3, constructing a multi-base station difference matrix based on visibility; Based on the difference values of all channel state information data collection times i, j, the difference metric matrix of the base station for a time period is obtained , and the formula is: (10) Given a base station BS and a user equipment UE, at two discrete measurement times i, j, the user equipment UE is visible to the base station if and only if the user equipment UE is visible to the base station , otherwise ; Based on weighted average combination A global difference metric matrix is obtained from the elements in the difference metric matrix of each base station BS to the user equipment UE A global difference metric matrix is obtained from the elements in the difference metric matrix of each base station BS to the user equipment UE , the formula is: (11) ， ， wherein, is a channel state information data acquisition time i,j difference value based on the training model output; is a channel state information data acquisition time j,i difference value based on the training model output; is a base station difference value of channel state information data at two discrete measurement times i,j with a user equipment UE; is a user equipment UE visibility at two discrete measurement times i,j with a base station ; is a time domain tensor of channel state information data at acquisition times i,j , in a high dimensional space . 5.The global positioning method of constructing channel graph based on deep learning difference matrix metric according to claim 4, characterized in that, Step 3 specifically comprises: A time domain tensor of time domain channel state data of any two discrete collection times i, j is obtained by using a multi-dimensional scaling manifold learning method to obtain mapping positions of both in a reduced dimension space , , namely: (12) The mapping calculation satisfies the limited condition: (13) wherein, is a time domain tensor , the distance between the mapped locations in the corresponding reduced dimension space , , denotes the maximum degree of approximation; For all Discrete-time time-domain channel state data, Midpoint-to-distance Corresponding The distance between them should satisfy the minimum mean square distance, that is: (14) All eligible constituting a channel state data channel graph in a reduced dimension space That is: (15)。 6.The global positioning method of constructing channel graph based on deep learning difference matrix metric according to claim 5, characterized in that, based on all the condition-satisfied constructing a channel state data channel graph in the reduced dimension space, the specific steps are: (1) using the element data in to calculate , , ; (2) Compute the inner product and further get the inner product matrix after dimension reduction where, is the mean of all elements in the th row of is the mean of all elements in the th column of is the global mean of ;​​ (3) Eigenvalue decomposition of the matrix where denotes the transpose of the eigenvector matrix , and is a diagonal matrix of eigenvalues, ; when there are 2 nonzero eigenvalues, , the time-domain channel state data channel diagram is obtained .​ 7.The global positioning method of constructing channel graph based on deep learning difference matrix metric according to claim 6, characterized in that, Step 4 specifically comprises: Step 4-1, when the base station collects and the user equipment collects the channel state data at L discrete times, the reference signal received power data set is synchronously acquired, and the position of the base station in the channel graph is estimated; Step 4-2, given the map space position of the center translation of each base station, the rotation angle of the base station compared with the estimated position in the channel graph is calculated, and the circular average method is used to estimate the global rotation angle; Step 4-3, using elliptical grid mapping for conformal transformation, mapping the circle to the square to eliminate the circularity in the channel graph; Step 4-4, constructing a homogeneous coordinate matrix, obtaining an affine transformation matrix based on a linear equation set, and solving; Step 4-5, global positioning based on the affine matrix.

Images