Massive MIMO-OFDM triple beam domain channel fingerprint construction and joint positioning and orientation method
By constructing a triple beam domain channel fingerprint and using an attention-based positioning and detection model, the problem of low accuracy in complex environments of wireless positioning technology is solved, and high-resolution simultaneous positioning and orientation sensing is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2025-12-18
- Publication Date
- 2026-06-23
Smart Images
Figure CN121619653B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless communication technology and relates to a method for constructing channel fingerprints in the triple beam domain of large-scale MIMO-OFDM and for joint localization and orientation. Background Technology
[0002] With the development of information technology, location-based services are increasingly being used in the Industrial Internet of Things (IIoT) and autonomous driving fields. However, the performance of the widely adopted Global Positioning System (GPS) degrades significantly in obstacle-filled urban or indoor environments. On the other hand, wireless networks, in addition to their primary communication functions, also have inherent advantages in positioning and sensing.
[0003] Wireless positioning systems estimate location based on reference signals between base stations and user terminals, primarily falling into two categories: geometry-based methods and fingerprint-based methods. Geometry-based methods utilize trigonometric properties to estimate location using parameters such as received signal strength, time of arrival (TOA), time difference of arrival (TDOA), and direction of arrival. Fingerprint-based methods uniquely map certain features of the reference signal to the user terminal coordinates, determining the user terminal's location by matching the measured fingerprint with a fingerprint database. In complex non-line-of-sight environments, multipath information significantly degrades the performance of geometry-based positioning methods, but provides richer fingerprints. Received signal strength is a commonly used fingerprint, but it reflects only limited channel information and suffers from low discrimination and poor stability. Channel State Information (CSI) contains multipath information in both the angle and time delay domains, generating fingerprints with higher discrimination and thus improving positioning accuracy. As a key technology for 5G and next-generation communication systems, Massive Multiple-Input Multiple-Output (MIMO) Orthogonal Frequency Division Multiplexing (OFDM) can capture high-resolution angle-delay domain multipath information, achieving high-precision positioning. Current CSI fingerprinting is limited to static estimation of the user terminal's position. For dynamic information such as motion state, it relies heavily on inertial measurement units or continuous CSI recordings, but these methods usually require additional hardware or complex processing.
[0004] Besides the inherent characteristics of fingerprints, the storage and matching methods of fingerprint databases are also key factors affecting wireless positioning performance. If the fingerprint matrix (or tensor) is considered as a single-channel (or multi-channel) image, and the corresponding coordinates as labels, then the matching problem between them can be viewed as a computer vision task. Therefore, many studies have used deep learning for fingerprint localization. Early studies used simple clustering algorithms, such as weighted k-nearest neighbors, to estimate the location. Recent studies focus more on fingerprint features, adjusting neural networks according to fingerprint structure, and designing more efficient localization methods, such as deep convolutional neural networks, 3D convolutional neural networks, and cooperative networks based on autoencoders. However, these methods fail to consider the impact of fingerprint sparsity on network structure. Traditional convolutional neural networks rely on the local correlation of data, while sparse fingerprints destroy spatial continuity, causing training results to converge only to local optima. Therefore, for the needs of next-generation mobile communication technologies, adopting accurate and efficient localization and orientation sensing algorithms has become a key issue in the current technological development. Summary of the Invention
[0005] Objective: To address the shortcomings of existing technologies for large-scale MIMO-OFDM systems, this invention aims to propose a triple beam domain channel fingerprint construction method. This method achieves high-resolution fingerprints supporting simultaneous localization and orientation while maintaining a lightweight fingerprint structure. Furthermore, this invention proposes a joint localization and orientation method based on beam domain channel characteristics to improve the system's localization and motion direction estimation accuracy, thus meeting the demands of next-generation mobile communication technologies.
[0006] Technical Solution: To achieve the above-mentioned objectives, the present invention provides the following technical solution:
[0007] In a first aspect, the present invention provides a method for constructing a triple beam domain channel fingerprint, comprising the following steps:
[0008] In a large-scale MIMO-OFDM system, the spatial-frequency-time domain channel between the user and the base station is transformed into a triple beam domain. The base station uses a uniform planar array, and the elevation and azimuth angles of the received signal arrival direction are obtained based on the uniform planar array. The spatial-frequency-time domain channel tensor is obtained by performing modulo operations with the triple beam domain channel response tensor and the angle, time delay, and Doppler domain discrete Fourier transform matrices, respectively.
[0009] Constructing triple beam domain channel fingerprints based on the modulus-square tensor of the triple beam domain channel; where the modulus-square tensor refers to the Hadamard product of a tensor and its conjugate tensor.
[0010] Furthermore, based on the regular grid arrangement characteristics of the uniform planar array in the horizontal and vertical dimensions, the array response vectors of each row and column of antennas are independent of each other, representing the horizontal spatial domain and the vertical spatial domain, respectively; the spatial domain array response vector of the triple beam domain channel is obtained by the Kronecker product of the horizontal and vertical spatial domain array response vectors.
[0011] Furthermore, the uniform planar array undergoes discrete Fourier transforms on its array responses in the horizontal and vertical dimensions, respectively. The angle domain transformation matrix is obtained by the Kronecker product of the horizontal and vertical transformations. The time delay domain transformation matrix is a submatrix of the standard discrete Fourier transform matrix. The Doppler domain transformation matrix is a submatrix of the discrete Fourier transform matrix with time offset.
[0012] Secondly, the present invention provides a joint localization and orientation estimation method, comprising the following steps:
[0013] Using the user terminal's channel, a triple beam domain channel fingerprint is generated according to the method described in the first aspect;
[0014] The angle-time delay domain information and the Doppler domain information are separated, and the position mask matrix is generated using the angle-time delay domain matrix. The angle-time delay domain matrix and the position mask matrix are processed by the mask-enhanced attention mechanism-based localization detection model to estimate the three-dimensional position coordinates of the user terminal.
[0015] By using a fusion-enhanced attention-based classification model to process Doppler domain information and estimated 3D position coordinates, the motion direction of the user terminal can be estimated.
[0016] Furthermore, the angle-delay domain information is a sparse matrix obtained by normalizing and summing the triple beam domain channel fingerprint along the Doppler domain; the position mask matrix is a matrix obtained by binarizing the angle-delay domain information based on a threshold; and the Doppler domain information is a sparse vector obtained by normalizing and summing the triple beam domain channel fingerprint along the angle and delay domains.
[0017] Furthermore, the mask-enhanced attention-based localization and detection model employs a mask-enhanced detection Transformer model, including a backbone network, a detection Transformer encoder / decoder, and an output layer. The model input is an angle-time-delay domain matrix and a corresponding position mask matrix, and the output is a three-dimensional coordinate vector. The backbone network uses either an input-level early stitching fusion method or a feature-level dual embedding fusion method. The input-level early stitching fusion method first stitches the angle-time-delay domain matrix and the position mask matrix in the time-delay domain dimension, and then extracts deep features through a nonlinear projection module. The feature-level dual embedding fusion method first processes the angle-time-delay domain matrix and the position mask matrix separately through an embedding layer, and then performs stitching and deep feature extraction. The nonlinear projection module consists of a linear layer, a ReLU activation function, and a normalization layer. The embedding layer has the same structure as the nonlinear projection module, but uses optimized hyperparameter configurations.
[0018] Furthermore, the fusion-enhanced attention-based classification model employs a fusion-enhanced Transformer classifier, including a fusion module, a Transformer encoder / decoder, and a multi-classifier. The model input consists of the estimated three-dimensional coordinate vector and Doppler domain vector from the user terminal, and the output is a scalar value representing direction. The fusion module has a dual-input structure, where the Doppler domain vector and the three-dimensional coordinate vector are concatenated and output after passing through a linear layer, a normalization layer, and a GELU activation function, respectively. The multi-classifier is a feedforward neural network containing two fully connected layers. Each fully connected layer is followed by a GELU activation function, layer normalization, and Dropout operation, and finally, the classification result is output through a Softmax layer.
[0019] Furthermore, the mask-enhanced attention-based localization and detection model is trained using localization mean square error and... The regularization term is used as the loss function, and labeled supervised training is used; the fusion-enhanced attention-based classification model uses cross-entropy and label smoothing as loss functions during training, and labeled supervised training is used.
[0020] Thirdly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the triple beam domain channel fingerprint construction method described in the first aspect or the joint positioning and orientation estimation method described in the second aspect.
[0021] Fourthly, the present invention provides a large-scale MIMO-OFDM communication system, including a base station and multiple user terminals. The user terminals are used to cooperate with the base station to generate a triple beam domain channel fingerprint. The base station includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the triple beam domain channel fingerprint construction method described in the first aspect or the joint positioning and orientation estimation method described in the second aspect.
[0022] Beneficial Effects: Compared with existing technologies, this invention has the following advantages: 1. This invention proposes a triple-beam domain channel fingerprint construction method, using a uniform planar array to make the channel fingerprint contain three-dimensional position information, Doppler domain information to support orientation estimation, and utilizing channel sparsity to ensure lightweight fingerprint structure, resulting in a high-resolution fingerprint that supports simultaneous localization and orientation. 2. The mask-enhanced attention-based localization and detection model proposed in this invention fully utilizes the mask mechanism and fingerprint sparsity, significantly improving position estimation performance while accelerating model training. 3. The fusion-enhanced attention-based classification model proposed in this invention significantly improves the robustness and accuracy of user terminal motion direction estimation through global spatiotemporal attention and multi-feature fusion. Attached Figure Description
[0023] Figure 1 This is a flowchart of the triple beam domain channel fingerprint construction method according to an embodiment of the present invention.
[0024] Figure 2 This is a flowchart of a large-scale MIMO-OFDM joint localization and orientation method based on beam domain channel characteristics, according to an embodiment of the present invention.
[0025] Figure 3 This is a schematic diagram of a large-scale MIMO-OFDM system and its frame structure in an embodiment of the present invention.
[0026] Figure 4 This is a diagram of the positioning and orientation sensing network architecture in an embodiment of the present invention.
[0027] Figure 5 This is a structural diagram of the mask-enhanced detection Transformer regression model in an embodiment of the present invention.
[0028] Figure 6 This is a structural diagram of the enhanced Transformer orientation classifier in an embodiment of the present invention.
[0029] Figure 7 This is a graph showing the cumulative distribution function of the positioning error in an embodiment of the present invention.
[0030] Figure 8 This is a diagram showing the confusion matrix results of the directional estimation in an embodiment of the present invention. Detailed Implementation
[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0032] like Figure 1 As shown, the triple beam domain channel fingerprint construction method disclosed in this embodiment of the invention includes: transforming the spatial-frequency-time domain channel between the user and the base station into a triple beam domain in a large-scale MIMO-OFDM system; wherein the base station uses a uniform planar array, and obtains the elevation angle and azimuth angle of the direction of arrival of the received signal based on the uniform planar array; the spatial-frequency-time domain channel tensor is obtained by performing modulo product operations with the triple beam domain channel response tensor and the angle, time delay, and Doppler domain discrete Fourier transform matrices respectively; constructing a triple beam domain channel fingerprint based on the modulo-squared tensor of the triple beam domain channel; wherein the modulo-squared tensor refers to the Hadamard product of a tensor and its conjugate tensor.
[0033] In this embodiment, the space-frequency-time domain channel includes the elevation angle, azimuth angle, time delay, and Doppler frequency information of the base station received signal, which are represented by a three-dimensional tensor; the triple beam domain channel specifically refers to the angle-time delay-Doppler domain, which can be obtained by performing discrete Fourier transform along each dimension of the space-frequency-time domain channel.
[0034] Specifically, based on the regular grid arrangement characteristics of the uniform planar array in the horizontal and vertical dimensions, the array response vectors of each row and column of antennas are independent of each other, representing the horizontal spatial domain and the vertical spatial domain, respectively; the spatial domain array response vector of the triple beam domain channel is obtained by the Kronecker product of the horizontal and vertical spatial domain array response vectors.
[0035] The uniform planar array undergoes Discrete Fourier Transform (DFT) on its horizontal and vertical array responses, respectively. The Kronecker product of the horizontal and vertical transforms yields the angular domain transform matrix. The time-delay domain transform matrix is a submatrix of the standard Discrete Fourier Transform matrix. The Doppler domain transform matrix is a submatrix of the Discrete Fourier Transform matrix with time offset.
[0036] In this embodiment, the modulus-square tensor of the triple beam domain channel has a first dimension representing the angle domain power containing arrival angle information, a second dimension representing the time delay domain power containing arrival time information, and a third dimension representing the Doppler domain power containing motion direction information.
[0037] Based on the scheme of this embodiment, the spatial-frequency-time domain channel is smaller in size after being transformed into a triple beam domain channel, and has sparsity in the angle-delay domain and block sparsity in the Doppler domain; the triple beam domain channel fingerprint can be regarded as a channel power tensor, which is a stable fingerprint containing multipath information and has the same sparsity as the triple beam domain channel tensor; the constructed triple beam domain channel fingerprint contains a variety of channel information, including the received signal strength, angle of arrival, and time of arrival for positioning, and the Doppler frequency for estimating motion state.
[0038] The sparsity of triple beam domain channel fingerprints is related to the multipath signal angle spread, cyclic prefix length, and number of time slots per frame. The positioning performance of triple beam domain channel fingerprints is related to the configuration of large-scale MIMO-OFDM systems. The larger the number of antennas, subcarriers, and OFDM symbols, the more complete the effective positioning information in the space-frequency-time domain is retained after transformation to the triple beam domain, resulting in higher fingerprint resolution and more significant sparsity.
[0039] Appendix Figure 2 A simplified flowchart of the large-scale MIMO-OFDM joint localization and orientation method based on beam domain channel characteristics disclosed in this invention is provided. Specifically, the large-scale MIMO-OFDM joint localization and orientation method based on beam domain channel characteristics includes: the base station generating a corresponding triple beam domain channel fingerprint using the user terminal's channel; the base station separating the angle-delay domain information from the Doppler domain information and generating a position mask matrix using the angle-delay domain matrix; the base station processing the angle-delay domain matrix and the position mask matrix using a mask-enhanced attention-based localization detection model to estimate the three-dimensional position coordinates of the user terminal; and the base station processing the Doppler domain information and the estimated three-dimensional position coordinates using a fusion-enhanced attention-based classification model to estimate the motion direction of the user terminal.
[0040] Specifically, the angle-delay domain information is a sparse matrix obtained by normalizing and summing the channel fingerprint of the triple beam domain along the Doppler domain; the position mask matrix is a matrix after binarizing the angle-delay domain information based on a threshold, with elements greater than the threshold set to 1 and the rest set to 0; the Doppler domain information is a sparse vector obtained by normalizing and summing the channel fingerprint of the triple beam domain along the angle domain and the time delay domain.
[0041] In one possible implementation, the mask-enhanced attention-based localization detection model employs a mask-enhanced detection Transformer model, including a backbone network, a detection Transformer encoder / decoder, and an output layer. The model input is an angle-time-delay domain matrix and a corresponding position mask matrix, and the output is a three-dimensional coordinate vector. The backbone network uses either an input-level early stitching fusion method or a feature-level dual embedding fusion method. The input-level early stitching fusion method first stitches the angle-time-delay domain matrix and the position mask matrix in the time-delay domain dimension, and then extracts deep features through a nonlinear projection module. The feature-level dual embedding fusion method first processes the angle-time-delay domain matrix and the position mask matrix separately through an embedding layer, and then performs stitching and deep feature extraction. The nonlinear projection module consists of a linear layer, a ReLU activation function, and a normalization layer. The embedding layer has the same structure as the nonlinear projection module, but uses specially optimized hyperparameter configurations.
[0042] In one possible implementation, the fusion-enhanced attention-based classification model employs a fusion-enhanced Transformer classifier, including a fusion module, a Transformer encoder / decoder, and a multi-classifier. The model input consists of the estimated 3D coordinate vector and Doppler domain vector from the user terminal, and the output is a scalar value representing direction. The fusion module has a dual-input structure, where the Doppler domain vector and the 3D coordinate vector are concatenated and output after passing through a linear layer, a normalization layer, and a GELU activation function, respectively. The multi-classifier is a feedforward neural network containing two fully connected layers. Each fully connected layer is followed by a GELU activation function, layer normalization, and Dropout operation, and finally, the classification result is output through a Softmax layer.
[0043] In one possible implementation, the mask-enhanced attention-based localization detection model is trained using the localization mean square error and... The regularization term is used as the loss function, and labeled supervised training is used; the fusion-enhanced attention-based classification model uses cross-entropy and label smoothing as loss functions during training, and labeled supervised training is used.
[0044] The following detailed explanation of the implementation process of the large-scale MIMO-OFDM joint localization and orientation method based on beam domain channel characteristics, as disclosed in this invention, is based on a specific communication system example. It should be noted that the method of this invention is applicable not only to the specific system model illustrated in the example below, but also to system models with other configurations.
[0045] I. System Configuration and Channel Model
[0046] Appendix Figure 3This invention provides a schematic diagram of a massive MIMO-OFDM system and its frame structure as disclosed in an embodiment of the present invention. Consider a single-cell time-division duplex massive MIMO-OFDM system, which includes a centrally located base station and... Individual user terminals. Using a collection. This represents all user terminals. The base station is equipped with a uniform planar array positioned in the XZ plane, with each row containing... There are 1 antenna, with 1 in each column. There are [number] antennas, and the total number of antennas is [number]. The spacing between adjacent antennas in the same row and column are as follows: and Each user terminal is equipped with a single antenna, which is randomly distributed within the cell.
[0047] OFDM modulation is used, and the number of subcarriers is The sampling interval length is Set the loop prefix length to Then the subcarrier spacing Length of each OFDM symbol The carrier wavelength is In the temporal domain, each frame contains Each time slot is divided into 1 time slot. Each frame contains OFDM symbols. That is, each frame contains OFDM symbols. For convenience, all variable indices and dimensions start from 0, including the current time slot and the previous one. Multiple time slots are merged into one frame. Considering the uplink scenario, the signal transmitted by each user terminal reaches the base station through multiple paths. First, we focus on the scenario within a single OFDM symbol. The signal sent by each user terminal passes through After propagating along a path (including reflection and scattering), the signal is received by the base station. Using a set... Let represent all paths. These paths are assumed to be generalizedly stationary and independent, and may include line-of-sight propagation conditions or be entirely non-line-of-sight paths. Let the _i_ be the _j_i ... The angle of arrival (AHA) of the path is decomposed into vertical pitch components. and the azimuth component in the horizontal direction The array response vectors in the horizontal and vertical spatial domains can be represented as follows:
[0048] (1)
[0049] (2)
[0050] in, It is the imaginary unit. This indicates the transpose operation.
[0051] The spatial domain array response vector of UPA is obtained by the Kronecker product of the horizontal and vertical spatial domain array response vectors, which can be specifically expressed as:
[0052] (3)
[0053] in Indicates the Kronecker product. Indicates the first OK( (represents rounding down) List( express right (Modular operation) The phase factor of the antenna. The first... The first user terminal The Doppler frequency of the path is denoted as Its phase factor is . No. The delay from each path to the base station is expressed as: During a single OFDM symbol, the first The channel response vector of a user terminal can be expressed as:
[0054] (4)
[0055] in Indicates the first The user terminal in the first The gain along each path follows a variance of . complex Gaussian distribution , Indicates time delay. This represents the Dirac function. In the context of... In the frame structure of an OFDM symbol, it is assumed that the channel state changes between symbols but remains constant within a single symbol. OFDM symbols (i.e.) The phase factor on ) is
[0056] (5)
[0057] Assuming the duration of the loop prefix exceeds the latency of all user terminal paths, i.e., satisfying... . No. The phase factor on each subcarrier is expressed as follows:
[0058] (6)
[0059] For the The user terminal, the first In the nth orthogonal frequency division multiplexing symbol The channel response vector on each subcarrier can be represented as follows:
[0060] (7)
[0061] The channel model in this embodiment of the invention integrates information from the spatial domain, time domain, and frequency domain, respectively corresponding to... , and Compared to common space-frequency-time domain channel models, this invention uses a uniform planar array to extend the spatial domain from two dimensions to three dimensions, thereby enriching the available fingerprint information. Frequency domain steering vector. and time-domain steering vector Represented as
[0062] (8)
[0063] (9)
[0064] in This is the first OFDM symbol in this frame. This also determines the range of multipath delay and Doppler frequency:
[0065]
[0066] With these three guiding vectors, the first... The space-frequency-time domain tensor channel model of a user terminal is represented in the following form.
[0067] (10)
[0068] Tensor and outer product operation Defined as
[0069] (11)
[0070] tensor It contains rich channel information. However, the space-frequency-time domain channel tensor contains a large number of elements, leading to a high computational load. To address this issue, the space-frequency-time domain channel tensor is transformed into the triple-beam (TB) domain, and the sparse TB domain channel tensor is used to handle the localization problem. The Discrete Fourier Transform (DFT) matrices of the three beam domains are used:
[0071] (12)
[0072] (13)
[0073] (14)
[0074] Where the matrix , yes The front of the point DFT matrix Submatrix, It has a time offset of The front of the point DFT matrix Example submatrix.
[0075] The transformation matrix in the angle domain is the Kronecker product of the transformation matrices of the uniform planar array in the horizontal and vertical dimensions, yielding the TB domain channel. The expression,
[0076] (15)
[0077] in Indicates the first The TB domain channel of the user terminal, its first The elements are From the above formula, we can see that the first... The SFT domain channel tensor of each user terminal is obtained by performing modulo operations with the TB domain channel response tensor on the angle, time delay, and Doppler domain discrete Fourier transform matrices, respectively. Specifically, it is obtained by performing a 1-modulus operation with the angle domain transform matrix, a 2-modulus operation with the time delay domain transform matrix, and a 3-modulus operation with the Doppler domain transform matrix. When performing... and of When multiplying by modulo, we get Each of its elements satisfies
[0078] (16)
[0079] Will Defined as a triple beam tensor, where each element is represented as follows:
[0080] (17)
[0081] For tensors and The Einstein product can be obtained And each of its elements satisfies
[0082] (18)
[0083] Multipath channels in wireless communication are influenced by the location of the user terminal and determined by the scattering environment between the user terminal and the base station. Changes in the scattering environment will produce different channel fingerprint characteristics. Tensor It contains rich channel information, including received signal strength, direction of arrival, time of arrival, and Doppler frequency. Different locations of the user terminal will generate different... Tensors are introduced to obtain more stable channel information. Second-order statistics:
[0084] (19)
[0085] This formula represents the space-frequency-time domain channel covariance tensor. It is a form of statistical CSI. It can effectively mitigate the effects of small-scale fading and maintain relatively stable temporal characteristics, thus making it suitable as a fingerprint feature. Due to the tensor... Due to its high dimensionality and computational complexity, the sparsity of the channel in the TB domain is utilized to simplify the calculation. The definition of Triple-Beam Fingerprint (TBF) is as follows:
[0086] (20)
[0087] in This represents the Hadamard product, and the TBF is also the channel power tensor in the TB domain. This type of TBF serves as a stable fingerprint feature, incorporating multiple channel characteristics. Notably, the TBF achieves positioning accuracy comparable to the SFTF while maintaining a more compact size.
[0088] II. Fingerprint Features
[0089] The accuracy of fingerprint localization is closely related to the distinguishability of the fingerprint. The TBF (Transient Frame Filter) incorporates various channel state information in the angle, time delay, and Doppler domains. Compared to fingerprints that rely on received signal strength, direction of arrival, or time of arrival, TBF offers higher resolution, enabling fingerprints generated by user terminals at different locations to have more significant distinguishability. For ease of description, the following definition is used: ,
[0090] (twenty one)
[0091] in yes The length, and For a large-scale MIMO-OFDM system, when all dimensions in the TB domain tend to infinity, the channel power of each path will be concentrated at a specific location, which can be expressed as:
[0092] (twenty two)
[0093] in, and It can be treated as an integer because and Smaller, and and Larger. And
[0094] (twenty three)
[0095] (twenty four)
[0096] when and When the value is large, it can be treated as an integer. If only one dimension in the time domain tends to infinity, then the channel power tensor will be concentrated at certain specific locations in that dimension.
[0097] In this embodiment of the invention, when , , and As it approaches infinity, the first The TBF of the first user terminal The channel energy of the multipath will be focused on the first path. Each angle and direction, time delay and the One OFDM symbol. In other words, Corresponding to the The direction of arrival, the first Arrival time and the first The average channel power is related to the Doppler frequency. This theorem clarifies the physical meaning of each element of the TBF, showing that the TBF is a unique fingerprint associated with its location and environment. The structure of the fingerprint tensor depends on the multipath information set. Because the channel environment differs at each location, the multipath information sets of two user terminals located at different locations are also different. Therefore, the corresponding TBF will also differ.
[0098] when , , and As the channel approaches infinity, the channel energy is concentrated only at specific locations in the TB domain, indicating that the TBF exhibits sparsity. Numerical simulations show that the TBF retains its sparsity even under large-scale finite-dimensional conditions.
[0099] This invention proposes a large-scale MIMO-OFDM tensor channel model and incorporates stable statistical information. As a spatial-frequency-time domain channel fingerprint. Therefore, It contains detailed information such as the received signal strength, direction of arrival, time of arrival, and Doppler frequency of each multipath component, making it a comprehensive and reliable fingerprint feature. To avoid the inherent defects caused by complex calculations, it is transformed to the TB domain to obtain a lightweight TBF. TBF, like the spatial-frequency-time domain channel fingerprint, retains key statistical information, and the collinearity relationship between the two fingerprint features can be used to measure their localization capabilities.
[0100] To characterize the similarity between fingerprints at different locations, a collinearity index for tensors is defined:
[0101] (25)
[0102] in This represents the Frobenius norm. Note the trace operation of tensors, which is explained below: For The conditions will be met. elements It is called a pseudo-diagonal element, and The trace is the sum of these pseudo-diagonal elements. Specifically,
[0103] (26)
[0104] Collinearity is used to measure the similarity of fingerprints at different locations. The lower the collinearity value, the higher the fingerprint's distinguishability.
[0105] In large-scale MIMO-OFDM systems, when , , and As the frequency-time domain channel fingerprint approaches infinity, the TBF (Transient-Temporal-Finger-Finger-Finger-Finger-Finger) exhibits equivalent collinearity, indicating that their localization capabilities are identical. Specifically, when... , , and When large enough, and The following relationship must be satisfied:
[0106] (27)
[0107] In this embodiment of the invention, when , , and As the frequency-frequency-time domain channel fingerprint approaches infinity, its collinearity is asymptotically equivalent to that of the TBF, and in... , , and When sufficiently large, they are approximately equal, possessing equivalent location information representation capabilities and exhibiting the same localization discrimination performance. However, TBF has advantages such as smaller size and superior sparsity; therefore, this invention selects TBF as the input to the network model.
[0108] III. Network Model
[0109] Based on the comprehensive environmental information contained in the TBF (Terrain-Based Function), this invention proposes a Localization and Orientation Awareness Network (LOA-Net) to simultaneously estimate the user terminal's position and motion direction. Given practical constraints, angle-delay domain information is chosen for localization, while Doppler domain information is dedicated to estimating motion direction. Accordingly, the localization function of LOA-Net is implemented through a mask-augmented detection transformer for regression (MaskDETR-Reg), while motion direction estimation is performed by a fusion-enhanced transformer for direction classification (Fusion-TDC). Appendix Figure 4 The overall architecture of LOA-Net is shown.
[0110] After data preprocessing, the angle-time delay domain and Doppler domain information are separated, and a position mask matrix is generated simultaneously. The angle-time delay domain data and the mask are aggregated into a batch input MaskDETR-Reg, while the Doppler domain information is processed in batches in the same way and input into Fusion-TDC.
[0111] Since the duration of a single frame is extremely short, assuming the user terminal is located at... It remains constant across each Doppler frequency dimension, i.e., satisfies Here Indicates the first The and the first The Doppler frequency correlation coefficient for each time slot. This assumption has been verified through numerical simulation. Define the... From the perspective of each user terminal, the latency domain data is as follows: Then we have:
[0112] (28)
[0113] Given each Having highly similar sparse patterns, Dimensionality compression was achieved while preserving all Doppler frequencies in the angle-delay domain. Normalization has been completed, and after batch aggregation, it can be directly input into MaskDETR-Reg.
[0114] Masks are used to constrain the regions of interest in a network, thereby reducing the impact of irrelevant regions on the output. Define the... The mask matrix of each user terminal is Then we have:
[0115] (29)
[0116] in, It is a threshold parameter; the larger its value, the smaller the range of the focused area.
[0117] Similar to The generation method, from Extracted Doppler domain information It is a one-dimensional vector, and its acquisition process is the same as above:
[0118] (30)
[0119] With user terminal location A bijective correspondence exists. From a computer vision perspective... It can be represented as a single-channel grayscale image, while the estimated location... This forms a continuous value vector. The Transformer architecture demonstrates significant advantages by overcoming the limitations of the limited receptive field in traditional CNNs. Utilizing self-attention and positional encoding, the Transformer can effectively model the correlations between multiple fingerprint features while capturing long-range dependencies. The detection Transformer adds a multi-query design to the classic Transformer, enabling the extraction of positional information from multiple perspectives. Based on this, this invention proposes MaskDETR-Reg for regression tasks. This network model can be viewed as a nonlinear function. ,Right now The specific architecture is attached. Figure 5 As shown.
[0120] This model employs a composite loss function. Combining the positioning mean square error and Regularization terms to optimize The total loss function is expressed as:
[0121] (31)
[0122] in The number of training samples. express Norm, To prevent overfitting Regularization term, These are regularization weight coefficients. These are the trainable parameters of the network.
[0123] Angle-delay domain information and position mask matrix in batch size Feature extraction is performed via a backbone network. Two fusion strategies are proposed: early concatenation fusion at the input layer and dual embedding fusion at the feature layer. These strategies effectively combine location information and focal region features. In early concatenation fusion, the input data and its corresponding mask matrix are concatenated along the temporal domain. Although input layer fusion ensures fast convergence, it lacks deep feature interaction. For dual embedding fusion, the input undergoes embedding, concatenation, and compression processing respectively. This represents the hidden layer dimension of the Transformer. The architectural difference between the early-concat fusion MaskDETR (EC-MaskDETR) and the dual-embedding fusion MaskDETR (DE-MaskDETR) lies only in the backbone network; the remaining components remain the same.
[0124] The features extracted by the backbone network are input into the detection Transformer encoder and decoder. (See attached...) Figure 5 middle, , , These represent the query vector, key vector, and value vector in the Transformer structure, respectively. and These represent the number of encoder and decoder layers, respectively. This represents the number of object queries. The multi-head self-attention mechanism and feedforward network employ the classic Transformer architecture. The fusion of positional encoding and self-attention enables the model to capture the intrinsic structure of non-sparse regions of the TBF and utilize this information in a focused manner. In the network of this invention, object queries represent... Parallel estimation results focusing on different input features. The output layer then... 3D features are mapped to 3D position coordinates, and then... The average of the coordinates is used as the final result to improve the robustness of the estimation.
[0125] The Doppler frequency is influenced by multiple factors, and the Doppler domain information in the TBF varies with position, velocity, and direction of motion. Using the coordinates estimated by MaskDETR-Reg as prior information, it is assumed that all user terminals move at the same speed and from... Deducing the terminal's motion direction. This invention constructs this problem as a multi-classification task and proposes Fusion-TDC, simultaneously processing the user terminal's spatial location and Doppler temporal sequence, and utilizing a global spatiotemporal attention mechanism to improve estimation accuracy. The angle range... Divided into 16 categories: . The direction is Its unique hot code is Fusion-TDC is considered a nonlinear function. , making Appendix Figure 6 The specific architecture of Fusion-TDC is shown.
[0126] loss function Combining cross-entropy loss and label smoothing loss, it is expressed as:
[0127] (32)
[0128] in
[0129] (33)
[0130] in The total number of categories, As a smoothing factor, These are weighting coefficients. Adjust as needed. and It can effectively suppress overfitting.
[0131] In the fusion module, the two inputs are respectively processed through a linear layer, layer normalization, and the GELU activation function for feature projection, and then concatenated. Tensors. The Transformer encoder and decoder employ a classic architecture, with... Figure 6 middle and These represent the number of encoder and decoder layers, respectively. Finally, the decoded tensor is reconstructed into a matrix, which is then used by a classifier and the Softmax function to generate a classification output.
[0132] IV. Implementation Results
[0133] To enable those skilled in the art to better understand the present invention, the performance results of the channel information acquisition method in this embodiment under specific configurations are given below.
[0134] OFDM parameters are configured according to the 3GPP LTE standard, and the TBF (Base Station Frame) is generated based on the indoor non-line-of-sight scenario defined in 3GPP 38.901. The core simulation parameters are as follows: number of base station antennas... carrier frequency Subcarrier spacing , the number of subcarriers Circular prefix length Number of time slots per frame Number of symbols per time slot User terminal speed .
[0135] Base station height is Located in the center of the cell with the antenna plane perpendicular to the ground. The user terminal height is set to... , and To simulate different floors. The sampling area is centered on the base station for coverage. Scope, each floor TBFs are generated at intervals. For the regression task, the training set contains 2883 TBFs (per floor). The test set contains 1200 TBFs (400 per floor) randomly distributed within the sampling area. In the multi-class classification task, the training set contains 16 TBFs with different orientations, sampled at 2883 preset locations. The test set generates 9920 TBFs with random positions and orientations. MaskDETR-Reg and Fusion-TDC are trained separately, with the coordinate input of Fusion-TDC replaced by the real labels. The training scheme employs adaptive learning rate scheduling and an early stopping strategy, triggering termination when the validation loss fails to improve for 100 consecutive rounds.
[0136] The dataset TBF used for this implementation validation was generated and saved using MATLAB R2023a, and the training and testing processes were completed using PyTorch 2.6. The simulation workstation ran on a platform equipped with a 12th generation Intel Core i7-12700 processor and an NVIDIA GeForce RTX 3060 graphics card.
[0137] To evaluate the performance of the proposed EC-MaskDETR and DE-MaskDETR, this implementation uses a CNN method as a benchmark. The localization error is measured by the difference between the estimated value and the true label. Norm calculation. For the two comparison methods (2D CNN and 3D CNN), only the kernel size of some convolutions in the convolution optimization module was adjusted to adapt to the input size, and the remaining hyperparameters were tuned to ensure that each network converged fully.
[0138] Appendix Figure 7 The cumulative distribution function of the localization error on the test set is shown. The detection-based Transformer network consistently outperforms the CNN-based network in localization accuracy. DE-MaskDETR achieves the highest accuracy, with an average error of [missing value]. and Error less than EC-MaskDETR performs slightly worse, with an average error of [missing information]. These results indicate that feature-level backbone networks can achieve deeper feature extraction.
[0139] Appendix Figure 8 The confusion matrix showing the test results is presented, with each cell displaying the number of samples (in parentheses) and the corresponding recall (the proportion of correctly identified true positives). The recall of the diagonal elements in the confusion matrix is significantly higher than that of the off-diagonal elements, indicating higher classification accuracy. Specifically, the number of correctly classified samples in the test set was 7865, achieving an accuracy of [missing information]. It is worth noting that in misclassification cases, categories closer to the diagonal typically exhibit higher recall, indicating that most incorrect estimates are still close to the correct classification.
[0140] This invention also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned triple beam domain channel fingerprint construction method or joint positioning and orientation estimation method.
[0141] The program code used to implement the method of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the steps of the method of the present invention are performed. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.
[0142] This invention also discloses a large-scale MIMO-OFDM communication system, including a base station and multiple user terminals. The user terminals are used to cooperate with the base station to generate a triple beam domain channel fingerprint. The base station includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the aforementioned triple beam domain channel fingerprint construction method or joint positioning and orientation estimation method. In this large-scale MIMO-OFDM communication system, the base station establishes a space-frequency-time domain channel model between itself and different users, and transforms it into a triple-beam domain channel model. The base station uses the user terminal's channel to generate a corresponding triple-beam domain channel fingerprint. The base station separates the angle-delay domain information from the Doppler domain information and uses the angle-delay domain matrix to generate a position mask matrix. The base station uses a mask-enhanced attention-based positioning and detection model to process the angle-delay domain matrix and the position mask matrix to estimate the three-dimensional position coordinates of the user terminal. The base station uses a fusion-enhanced attention-based classification model to process the Doppler domain information to estimate the motion direction of the user terminal. The frame length is small enough that the position of the user terminal remains unchanged within a single frame. The number of antennas, subcarriers, and OFDM symbols is large enough that the multipath channel power can be concentrated at a specific location.
[0143] In the embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways without departing from the spirit and scope of this application. The current embodiments are merely exemplary examples and should not be considered limiting, nor should the specific content given limit the purpose of this application. For example, some features may be omitted or not implemented.
[0144] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.
Claims
1. A method for constructing a triple beam domain channel fingerprint, characterized in that, Includes the following steps: In a large-scale MIMO-OFDM system, the spatial-frequency-time domain channel between the user and the base station is transformed into a triple beam domain. The base station uses a uniform planar array, and the elevation and azimuth angles of the received signal arrival direction are obtained based on the uniform planar array. The spatial-frequency-time domain channel tensor is obtained by performing modulo operations with the triple beam domain channel response tensor and the angle, time delay, and Doppler domain discrete Fourier transform matrices, respectively. Constructing triple beam domain channel fingerprints based on the modulus-square tensor of triple beam domain channels; The modulus square tensor refers to the Hadamard product of a tensor and its conjugate tensor. The triple beam domain channel refers to the angle-delay-Doppler domain, which is obtained by performing discrete Fourier transforms along each dimension of the space-frequency-time domain channel; The uniform planar array has regular grid arrangement characteristics in the horizontal and vertical dimensions. The array response vectors of each row and column of antennas are independent of each other, representing the horizontal spatial domain and the vertical spatial domain, respectively. The spatial domain array response vector of the triple beam domain channel is obtained by the Kronecker product of the horizontal and vertical spatial domain array response vectors. The uniform planar array undergoes Discrete Fourier Transform (DFT) on its horizontal and vertical array responses, respectively. The angle domain transformation matrix is obtained by the Kronecker product of the horizontal and vertical transformations. The time delay domain transformation matrix is a submatrix of the standard Discrete Fourier Transform matrix. The Doppler domain transformation matrix is a submatrix of the Discrete Fourier Transform matrix with time offset. The first dimension of the modulus-squared tensor of the triple beam domain channel represents the angle domain power containing arrival angle information, the second dimension represents the time delay domain power containing arrival time information, and the third dimension represents the Doppler domain power containing motion direction information.
2. A joint positioning and orientation estimation method, characterized in that, Includes the following steps: Using the channel of the user terminal, a triple beam domain channel fingerprint is generated according to the method described in claim 1; The angle-time delay domain information is separated from the Doppler domain information, and the position mask matrix is generated using the angle-time delay domain matrix. A mask-enhanced attention-based localization detection model is used to process the angle-delay domain matrix and the position mask matrix to estimate the three-dimensional position coordinates of the user terminal. By using a fusion-enhanced attention-based classification model to process Doppler domain information and estimated 3D position coordinates, the motion direction of the user terminal can be estimated.
3. The joint positioning and orientation estimation method according to claim 2, characterized in that: The angle-delay domain information is a sparse matrix, obtained by normalizing and summing the triple beam domain channel fingerprint along the Doppler domain; The position mask matrix is a matrix obtained by binarizing angle-delay domain information based on a threshold. The Doppler domain information is a sparse vector, obtained by normalizing and summing the triple beam domain channel fingerprint along the angle domain and time delay domain.
4. The joint positioning and orientation estimation method according to claim 2, characterized in that: The mask-enhanced attention-based localization and detection model employs a mask-enhanced detection Transformer model, including a backbone network, a detection Transformer encoder / decoder, and an output layer. The model input is an angle-time-delay domain matrix and a corresponding position mask matrix, and the output is a three-dimensional coordinate vector. The backbone network uses either an input-level early splicing fusion method or a feature-level dual embedding fusion method. The input-level early splicing fusion method first splices the angle-time-delay domain matrix and the position mask matrix in the time-delay domain dimension, and then extracts deep features through a nonlinear projection module. The feature-level dual embedding fusion method first processes the angle-delay domain matrix and the position mask matrix separately through the embedding layer, and then performs splicing and deep feature extraction; the nonlinear projection module consists of a linear layer, a ReLU activation function, and a normalization layer; the embedding layer has the same structure as the nonlinear projection module, but adopts different optimized hyperparameter configurations.
5. The joint localization and orientation estimation method according to claim 2, characterized in that: The fusion-enhanced attention-based classification model employs a fusion-enhanced Transformer classifier, including a fusion module, a Transformer encoder / decoder, and a multi-classifier. The model input consists of the estimated three-dimensional coordinate vector and Doppler domain vector of the user terminal, and the output is a scalar value representing the direction. The fusion module has a dual-input structure. The Plüss domain vector and the three-dimensional coordinate vector are concatenated and output after passing through a linear layer, a normalization layer, and a GELU activation function, respectively. The multi-classifier is a feedforward neural network containing two fully connected layers. Each fully connected layer is followed by a GELU activation function, layer normalization, and Dropout operation, and finally outputs the classification result through a Softmax layer.
6. The joint positioning and orientation estimation method according to claim 2, characterized in that: The mask-enhanced attention-based localization and detection model uses localization mean square error and regularization term as loss functions during training, and employs labeled supervised training; the fusion-enhanced attention-based classification model uses cross-entropy and label smoothing as loss functions during training, and employs labeled supervised training.
7. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-6.
8. A large-scale MIMO-OFDM communication system, comprising a base station and multiple user terminals, characterized in that: The user terminal is used to cooperate with the base station to generate a triple beam domain channel fingerprint. The base station includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1-6.