A multi-base station joint wireless positioning method and system based on self-attention mechanism

Abstract

The present invention discloses a multi-base station joint wireless positioning method and system based on a self-attention mechanism, which relates to the field of deep learning and wireless communication technology and includes the following steps: collecting channel observation data from distributed base stations to form a multi-base station joint channel matrix; performing Gaussian normalization on the various dimensions of the multi-base station joint channel matrix to obtain a normalized multi-base station joint channel matrix and performing multi-channel convolution to obtain a feature fusion matrix; performing self-attention value calculation on the feature fusion matrix to obtain a distributed attention coding matrix; constructing a long short-term memory recurrent neural network for the distributed attention coding matrix to decode the spatial position coordinates of the target to be sensed. The present invention innovatively proposes the ideas of distributed base station perception, multi-base station channel information fusion, and self-attention encoding and decoding, improving the channel feature extraction method in existing wireless positioning algorithms and improving the base station perception accuracy in large-scale wireless communication scenarios.

Description

Technical Field

[0001] The present invention relates to the field of deep learning and wireless communication technology, and in particular to positioning technology in a new generation of wireless communication systems. More specifically, it relates to a multi-base station joint wireless positioning method and system based on a self-attention mechanism. Background Art

[0002] Wireless positioning technology uses wireless signal characteristics (such as signal strength, propagation time, and angle of arrival) to infer the spatial location of a target object or user. As a core capability of next-generation intelligent wireless communication systems, wireless positioning has broad application value in intelligent traffic navigation, autonomous driving, industrial IoT device tracking, emergency rescue, smart city management, and other fields. Especially in complex urban environments, high-precision positioning is the foundation for enabling cutting-edge applications such as vehicle-road collaboration and augmented reality services.

[0003] Current mainstream wireless positioning technologies include satellite positioning (such as GPS), single-base station CSI deep learning positioning, multi-base station post-fusion positioning, and traditional geometric positioning methods (such as TDOA, AOA, and RSS). Satellite positioning, while offering wide-area coverage, suffers from a sharp drop in accuracy in densely populated cities due to signal obstruction and multipath effects. Single-base station CSI deep learning models, while offering high local accuracy, have limited coverage and significant performance degradation away from the base station. Multi-base station post-fusion technology, due to independent data processing, fails to fully exploit spatiotemporal correlations across base stations, resulting in low information utilization. Traditional geometric positioning methods, limited by the assumption of line-of-sight propagation, lack robustness in complex non-line-of-sight and multipath scenarios. These technologies all face key bottlenecks such as multipath interference, limited coverage, poor dynamic adaptability, and inefficient data fusion, hindering the demand for high-precision positioning in complex, large-scale outdoor scenarios.

[0004] The attention mechanism is a computational model that mimics human cognitive focusing, dynamically assigning weights to highlight key information in the data. Its core features include the ability to model long-range dependencies, adaptive feature selection, and multimodal information fusion. These characteristics make it particularly suitable for wireless positioning scenarios. Through the self-attention mechanism, it can effectively capture the temporal and spatial correlations implicit in the channel state information of multiple base stations, distinguish multipath interference from valid signals, and thus improve positioning accuracy in complex environments.

[0005] In large-scale outdoor wireless communication scenarios, multi-base station joint positioning introduces spatial diversity, comprehensively utilizing observation data from multiple base stations to expand positioning coverage and mitigate local environmental interference using multi-perspective signal characteristics. This is particularly true in scenarios such as dense urban canyons and highly dynamic transportation hubs, where multi-base station collaboration can significantly alleviate signal obstruction and enhance system robustness.

[0006] In summary, the existing technologies have the following problems: (1) Dense building reflections lead to complex signal propagation paths, making it difficult for traditional methods to distinguish between direct signals and multipath components; (2) The positioning accuracy of a single base station decays with distance and cannot meet the requirements of wide-area high precision; (3) Existing multi-base station solutions mostly use post-estimation fusion, which does not fully exploit the cross-base station correlation characteristics of the original CSI data.

[0007] Therefore, it is an urgent problem for those skilled in the art to propose a multi-base station joint wireless positioning method and system based on the self-attention mechanism to solve the difficulties existing in the existing technology. Summary of the Invention

[0008] In view of this, the present invention provides a multi-base station joint wireless positioning method and system based on the self-attention mechanism, which solves the problems existing in the background technology.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] A multi-base station joint wireless positioning method based on a self-attention mechanism includes the following steps:

[0011] S1. Collect channel observation data from distributed base stations to form a multi-base station joint channel matrix;

[0012] S2. Performing Gaussian normalization on each dimension of the multi-base station joint channel matrix to obtain a normalized multi-base station joint channel matrix;

[0013] S3. Perform multi-channel convolution on the normalized multi-base station joint channel matrix to obtain a feature fusion matrix;

[0014] S4. Calculate the self-attention value of the feature fusion matrix to obtain the distributed attention encoding matrix;

[0015] S5. Construct a long short-term memory recurrent neural network based on the distributed attention encoding matrix to decode the spatial position coordinates of the target to be perceived.

[0016] Optionally, the multi-base station joint channel matrix in S1 is a three-dimensional matrix consisting of the number of base stations, base station antennas, and channel subcarrier frequencies, and is constructed in the following steps:

[0017] S11. Assume that the number of antennas of each base station is M , the base station establishes a wireless connection with the user through orthogonal frequency division multiplexing technology, where the number of channel subcarriers in orthogonal frequency division multiplexing is N , then the user channel state matrix observed by a single base station is R Expressed as:

[0018]

[0019] Where: Indicates antenna m In the n The channel state response observed in the subcarriers, ;

[0020] S12. Assume that the number of base stations in the distributed base station communication service scenario is L , each base station repeats step S11 to construct a multi-base station joint channel matrix H It is expressed as follows:

[0021]

[0022] Where: Indicates the L The user channel state matrix observed by the base station.

[0023] Optionally, the Gaussian normalization in S2 is the normalization of the first dimension, which is also the base station normalization. The specific steps are:

[0024] S21. Solve the mean matrix U and variance matrix :

[0025]

[0026]

[0027] Where: represents the average value of the multi-base station joint channel matrix in the base station dimension, represents the variance of the multi-base station joint channel matrix in the base station dimension; L Indicates the number of base stations in the distributed base station communication service scenario, express H The index in the matrix is The element value of

[0028] S22. Perform Gaussian normalization operation:

[0029]

[0030] Where: H represents the multi-base station joint channel matrix, represents the normalized multi-base station joint channel matrix.

[0031] Optionally, the multi-channel convolutional network in S3 includes a two-layer structure. The first convolutional structure contains L channels, and the second convolution structure is a single channel;

[0032] The input of the multi-channel convolutional network is the three-dimensional normalized multi-base station joint channel matrix , the output is a two-dimensional feature fusion matrix ,matrix Y The calculation formula for each element in is:

[0033]

[0034] Where: represents the multi-channel convolution kernel, b represents the bias component of the convolutional network, M represents the number of antennas at each base station, N Indicates the number of channel subcarriers in OFDM; Represents the matrix dimension of the convolution result, Indicates the number of convolutions of the convolution kernel along the antenna dimension during a single convolution process. Indicates the number of convolutions of the convolution kernel along the subcarrier dimension during a single convolution process. express The index in the matrix is The element value of .

[0035] Optionally, in S4, the self-attention value of the feature fusion matrix is ​​calculated as follows:

[0036] The feature fusion matrix Y It is also used as the query matrix and key matrix in the attention mechanism, and the calculation formula is:

[0037]

[0038] Where: A represents the attention encoding matrix, softmax represents the probability conversion function, V represents a value vector, Represents the convolution result matrix of the previous step Y The second dimension value of .

[0039] Optionally, in S5, the spatial position coordinates of the target to be sensed are obtained in the following manner:

[0040] Construct a decoding network based on long short-term memory recurrent neural network and linear deep neural network;

[0041] The attention encoding matrix A The long short-term memory recurrent neural network is input to extract time series features. The time series features enter the linear deep neural network with a nonlinear activation function, so that the learned time series features are mapped to spatial positions and the spatial position coordinates of the target to be perceived are obtained.

[0042] A multi-base station joint wireless positioning system based on a self-attention mechanism, executing any of the multi-base station joint wireless positioning methods based on a self-attention mechanism described above, comprising:

[0043] The channel matrix construction module is used to collect channel observation data from distributed base stations and form a multi-base station joint channel matrix;

[0044] A data preprocessing module is used to perform Gaussian normalization on the data of each dimension of the multi-base station joint channel matrix to obtain a normalized multi-base station joint channel matrix;

[0045] The channel feature fusion module is used to perform multi-channel convolution on the normalized multi-base station joint channel matrix to obtain a feature fusion matrix;

[0046] The multi-head self-attention encoding module is used to calculate the self-attention value of the feature fusion matrix to obtain a distributed attention encoding matrix;

[0047] The recurrent neural network decoding module is used to construct a long short-term memory recurrent neural network for the distributed attention encoding matrix to decode the spatial position coordinates of the target to be perceived.

[0048] It can be seen from the above technical solution that compared with the prior art, the present invention discloses a multi-base station joint wireless positioning method and system based on the self-attention mechanism, which has the following beneficial effects:

[0049] This invention innovatively proposes the ideas of distributed base station perception, multi-base station channel information fusion and self-attention encoding and decoding, which improves the channel feature extraction method in existing wireless positioning algorithms, is conducive to high-precision wireless positioning in large-scale outdoor wireless communication scenarios, and provides basic support for cutting-edge applications such as intelligent traffic navigation and unmanned driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0051] Figure 1 The multi-base station joint wireless positioning scene graph based on the self-attention mechanism provided by the present invention;

[0052] Figure 2 This is a structural diagram of the multi-base station joint wireless positioning system based on the self-attention mechanism provided by the present invention;

[0053] Figure 3This is a performance comparison chart of the wireless positioning method provided by the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] Reference Figure 1 The embodiment of the present invention discloses a multi-base station joint wireless positioning method based on a self-attention mechanism, comprising the following steps:

[0056] S1. Collect channel observation data from distributed base stations to form a multi-base station joint channel matrix;

[0057] S2. Performing Gaussian normalization on each dimension of the multi-base station joint channel matrix to obtain a normalized multi-base station joint channel matrix;

[0058] S3. Perform multi-channel convolution on the normalized multi-base station joint channel matrix to obtain a feature fusion matrix;

[0059] S4. Calculate the self-attention value of the feature fusion matrix to obtain the distributed attention encoding matrix;

[0060] S5. Construct a long short-term memory recurrent neural network based on the distributed attention encoding matrix to decode the spatial position coordinates of the target to be perceived.

[0061] Furthermore, the multi-base station joint channel matrix in S1 is a three-dimensional matrix consisting of the number of base stations, base station antennas, and channel subcarrier frequencies, and its construction method includes the following steps:

[0062] S11. Assume that the number of antennas of each base station is M , the base station establishes a wireless connection with the user through orthogonal frequency division multiplexing technology, where the number of channel subcarriers in orthogonal frequency division multiplexing is N , then the user channel state matrix observed by a single base station is R Expressed as:

[0063]

[0064] Where: Indicates antenna m In the n The channel state response observed in the subcarriers, ;

[0065] S12. Assume that the number of base stations in the distributed base station communication service scenario is L , each base station repeats step S11 to construct a multi-base station joint channel matrix H It is expressed as follows:

[0066]

[0067] Where: Indicates the L The user channel state matrix observed by the base station.

[0068] Furthermore, the Gaussian normalization in S2 is the normalization of the first dimension, which is also the base station normalization. The specific steps are:

[0069] S21. Solve the mean matrix U and variance matrix :

[0070]

[0071]

[0072] Where: represents the average value of the multi-base station joint channel matrix in the base station dimension, represents the variance of the multi-base station joint channel matrix in the base station dimension; L Indicates the number of base stations in the distributed base station communication service scenario, express H The index in the matrix is The element value of

[0073] S22. Perform Gaussian normalization operation:

[0074]

[0075] Where: H represents the multi-base station joint channel matrix, represents the normalized multi-base station joint channel matrix, and division represents the division of corresponding matrix elements.

[0076] Furthermore, the multi-channel convolutional network in S3 consists of two layers. The first layer of convolutional structure contains L channels, and the second convolution structure is a single channel;

[0077] The input of the multi-channel convolutional network is the three-dimensional normalized multi-base station joint channel matrix , the output is a two-dimensional feature fusion matrix ,matrix Y The calculation formula for each element in is:

[0078]

[0079] Where: represents the multi-channel convolution kernel, b represents the bias component of the convolutional network, M represents the number of antennas at each base station, N Indicates the number of channel subcarriers in OFDM; Represents the matrix dimension of the convolution result, Indicates the number of convolutions of the convolution kernel along the antenna dimension during a single convolution process. Indicates the number of convolutions of the convolution kernel along the subcarrier dimension during a single convolution process. express The index in the matrix is The element value of .

[0080] Furthermore, in S4, the self-attention value of the feature fusion matrix is ​​calculated as follows:

[0081] The feature fusion matrix Y It is also used as the query matrix and key matrix in the attention mechanism, and the calculation formula is:

[0082]

[0083] Where: A represents the attention encoding matrix, softmax represents the probability conversion function, V represents a value vector, Represents the convolution result matrix of the previous step Y The second dimension value of .

[0084] Furthermore, in S5, the spatial position coordinates of the target to be sensed are obtained in the following manner:

[0085] Construct a decoding network based on long short-term memory recurrent neural network and linear deep neural network;

[0086] The attention encoding matrix A The long short-term memory recurrent neural network is input to extract time series features. The time series features enter the linear deep neural network with a nonlinear activation function, so that the learned time series features are mapped to spatial positions and the spatial position coordinates of the target to be perceived are obtained.

[0087] and Figure 1 Corresponding to the method described above, the embodiment of the present invention also provides a multi-base station joint wireless positioning system based on the self-attention mechanism, which is used to Figure 1 The specific implementation of the method in the embodiment of the present invention is a multi-base station joint wireless positioning system based on the self-attention mechanism, which can be applied to computer terminals or various mobile devices, such as Figure 2 As shown, specifically including:

[0088] The channel matrix construction module is used to collect channel observation data from distributed base stations and form a multi-base station joint channel matrix;

[0089] A data preprocessing module is used to perform Gaussian normalization on the data of each dimension of the multi-base station joint channel matrix to obtain a normalized multi-base station joint channel matrix;

[0090] The channel feature fusion module is used to perform multi-channel convolution on the normalized multi-base station joint channel matrix to obtain a feature fusion matrix;

[0091] The multi-head self-attention encoding module is used to calculate the self-attention value of the feature fusion matrix to obtain a distributed attention encoding matrix;

[0092] The recurrent neural network decoding module is used to construct a long short-term memory recurrent neural network for the distributed attention encoding matrix to decode the spatial position coordinates of the target to be perceived.

[0093] In a specific embodiment, a multi-base station joint wireless positioning method based on a self-attention mechanism is applied in an outdoor mobile communication scenario. This scenario involves a complex urban canyon environment, including densely populated buildings, transportation hubs, and other areas with significant signal obstruction and multipath effects. Distributed base stations observe channel state information of mobile users, and the method provided in this embodiment achieves high-precision positioning in various outdoor scenarios.

[0094] In this embodiment, the number of distributed base stations is six, each equipped with eight antennas and employing 64-subcarrier orthogonal frequency division multiplexing (OFDM). Multiple base stations simultaneously observe user channels, constructing a three-dimensional multi-base station joint channel matrix. To extract user locations, this example proposes a fusion-encoding-decoding neural network consisting of three modules. The fusion module consists of a two-layer convolutional neural network. The first convolution layer uses a 3×4 convolution kernel with 6 channels and a stride of 1, while the second layer uses a 3×4 convolution kernel with 1 channel and a stride of 1, fusing the distributed base station channel features. The encoding module employs a six-head self-attention mechanism to encode channel features. The decoding module uses a single-layer long short-term memory unit (LSTM) connected to a linear deep neural network. This design adheres to the uniformity of feature dimensions across various neural network interfaces. The linear deep neural network consists of a five-layer structure: an input layer with 60 neural units, a middle layer with 512 neural units, and an output layer with 2 neural units, corresponding to the user's two-dimensional location coordinates. Each layer is connected through a nonlinear activation layer to enhance the nonlinear expression ability of the model.

[0095] Figure 3This paper presents a performance comparison of this algorithm with comparison algorithms in the same scenario. Method 1 uses the MFCNet network structure proposed by Z. Chen, Z. Zhang, et al., Method 2 uses the ResNet18 structure, and Method 3 uses a linear deep neural network structure. Method 1 represents a typical single-base station model, while Methods 2 and 3 represent traditional multi-base station models. It can be seen that, with six base stations working together, the proposed method achieves decimeter-level positioning accuracy, with 50% of positioning errors below 0.25m, significantly outperforming both single-base station models and traditional multi-base station fusion solutions.

[0096] In summary, this embodiment proposes a multi-base station joint wireless positioning method and system based on the self-attention mechanism, which provides a feasible solution for high-precision wireless positioning applications in outdoor large-scale wireless communication scenarios, and provides basic support for cutting-edge applications such as intelligent traffic navigation and unmanned driving.

[0097] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0098] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-base station joint wireless positioning method based on self-attention mechanism, characterized in that: The following steps are involved: S1. Collect channel observation data from distributed base stations to form a multi-base station joint channel matrix; S2. Performing Gaussian normalization on each dimension of the multi-base station joint channel matrix to obtain a normalized multi-base station joint channel matrix; S3. Perform multi-channel convolution on the normalized multi-base station joint channel matrix to obtain a feature fusion matrix; S4. Calculate the self-attention value of the feature fusion matrix to obtain the distributed attention encoding matrix; S5. Construct a long short-term memory recurrent neural network based on the distributed attention encoding matrix to decode the spatial position coordinates of the target to be perceived.

2. The multi-base station joint wireless positioning method based on the self-attention mechanism according to claim 1 is characterized in that: The multi-base station joint channel matrix in S1 is a three-dimensional matrix consisting of the number of base stations, base station antennas, and channel subcarrier frequencies. Its construction includes the following steps: S11. Assume that the number of antennas of each base station is M , the base station establishes a wireless connection with the user through orthogonal frequency division multiplexing technology, where the number of channel subcarriers in orthogonal frequency division multiplexing is N , then the user channel state matrix observed by a single base station is R Expressed as: Where: Indicates antenna m In the n The channel state response observed in the subcarriers, ; S12. Assume that the number of base stations in the distributed base station communication service scenario is L , each base station repeats step S11 to construct a multi-base station joint channel matrix H It is expressed as follows: Where: Indicates the L The user channel state matrix observed by the base station.

3. The multi-base station joint wireless positioning method based on the self-attention mechanism according to claim 1 is characterized in that: The Gaussian normalization in S2 is the normalization of the first dimension, also known as base station normalization. The specific steps are: S21. Solve the mean matrix U and variance matrix : Where: represents the average value of the multi-base station joint channel matrix in the base station dimension, represents the variance of the multi-base station joint channel matrix in the base station dimension; L Indicates the number of base stations in the distributed base station communication service scenario, express H The index in the matrix is The element value of S22. Perform Gaussian normalization operation: Where: H represents the multi-base station joint channel matrix, represents the normalized multi-base station joint channel matrix.

4. The multi-base station joint wireless positioning method based on the self-attention mechanism according to claim 1, characterized in that: The multi-channel convolutional network in S3 consists of two layers. The first layer of convolutional structure contains L channels, and the second convolution structure is a single channel; The input of the multi-channel convolutional network is the three-dimensional normalized multi-base station joint channel matrix , the output is a two-dimensional feature fusion matrix ,matrix Y The calculation formula for each element in is: Where: represents the multi-channel convolution kernel, b represents the bias component of the convolutional network, M represents the number of antennas at each base station, N Indicates the number of channel subcarriers in OFDM; Represents the matrix dimension of the convolution result, Indicates the number of convolutions of the convolution kernel along the antenna dimension during a single convolution process. Indicates the number of convolutions of the convolution kernel along the subcarrier dimension during a single convolution process. express The index in the matrix is The element value of .

5. The multi-base station joint wireless positioning method based on the self-attention mechanism according to claim 1, characterized in that: In S4, the self-attention value of the feature fusion matrix is ​​calculated as follows: The feature fusion matrix Y It is also used as the query matrix and key matrix in the attention mechanism, and the calculation formula is: Where: A represents the attention encoding matrix, softmax represents the probability conversion function, V represents a value vector, Represents the convolution result matrix of the previous step Y The second dimension value of .

6. The multi-base station joint wireless positioning method based on the self-attention mechanism according to claim 1, characterized in that: In S5, the spatial position coordinates of the target to be sensed are obtained in the following manner: Construct a decoding network based on long short-term memory recurrent neural network and linear deep neural network; The attention encoding matrix A The long short-term memory recurrent neural network is input to extract time series features. The time series features enter the linear deep neural network with a nonlinear activation function, so that the learned time series features are mapped to spatial positions and the spatial position coordinates of the target to be perceived are obtained.

7. A multi-base station joint wireless positioning system based on self-attention mechanism, characterized in that: Executing the multi-base station joint wireless positioning method based on the self-attention mechanism according to any one of claims 1 to 6, comprising: The channel matrix construction module is used to collect channel observation data from distributed base stations and form a multi-base station joint channel matrix; A data preprocessing module is used to perform Gaussian normalization on the data of each dimension of the multi-base station joint channel matrix to obtain a normalized multi-base station joint channel matrix; The channel feature fusion module is used to perform multi-channel convolution on the normalized multi-base station joint channel matrix to obtain a feature fusion matrix; The multi-head self-attention encoding module is used to calculate the self-attention value of the feature fusion matrix to obtain a distributed attention encoding matrix; The recurrent neural network decoding module is used to construct a long short-term memory recurrent neural network for the distributed attention encoding matrix to decode the spatial position coordinates of the target to be perceived.

Images