Multi-base-station cooperative robust positioning method and system based on deep learning

Abstract

The invention discloses a multi-base-station cooperative robust positioning method and system based on deep learning, and relates to the technical field of wireless communication, and the method comprises the steps: forming a multi-base-station multi-time-slot joint channel matrix through collecting the change channel observation data of distributed base stations; shifting time slot data of the multi-base-station multi-time-slot joint channel matrix, simulating a synchronization error of a real scene, and obtaining a multi-base-station joint channel matrix under the condition of the synchronization error; performing synchronization error calibration processing on the multi-base-station joint channel matrix under the synchronization error condition to obtain a calibrated multi-base-station joint channel matrix; and carrying out user coordinate positioning by using the calibrated multi-base-station joint channel matrix and the established positioning network. The performance of the multi-base-station cooperative positioning system is remarkably improved while the problem of synchronization errors is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to a multi-base station cooperative robust positioning method and system based on deep learning. BACKGROUND

[0002] With the large-scale deployment of the fifth generation mobile communication (5G) technology and the gradual evolution of the sixth generation communication system (6G) standard, high-precision positioning technology has become a key enabling technology to support the development of an intelligent Internet of Things society. In typical scenarios such as industrial Internet of Things (IIoT), vehicle-to-road cooperative systems (V2X), intelligent factories, and unmanned drone group control, the demand for centimeter-level positioning accuracy is showing an exponential growth trend. Multi-base station cooperative positioning technology achieves spatio-temporal fusion of measurement data between base stations by constructing a distributed sensing network. Compared with the global navigation satellite system (GNSS), it exhibits significant technical advantages in non-line-of-sight (NLOS) and indoor multipath environments, and has become a research focus in the field of ubiquitous positioning.

[0003] However, this technology faces severe synchronization challenges in actual deployment. Physical layer limitations such as hardware inherent delay and clock drift caused by multipath propagation result in microsecond-level or even nanosecond-level clock bias between base stations, directly affecting positioning accuracy and system performance. In traditional multi-base station positioning methods, synchronization errors cause deviations in the measurement of signal time of arrival (TOA) or time difference of arrival (TDOA), which in turn causes positioning errors when using distributed algorithms for joint positioning.

[0004] To address the synchronization error problem in multi-base station cooperative positioning, it is necessary to further explore its influence mechanism and propose effective error compensation methods, which have important theoretical significance and practical application value. In multi-base station cooperative positioning in complex environments and high dynamic applications, synchronization error calibration is an important direction to improve positioning stability and anti-interference capability, and is also a key to promoting multi-base station cooperative positioning technology from theoretical models to practical deployment. However, current research on multi-base station cooperative positioning is mostly based on the ideal assumption of no synchronization error, so it is necessary to fill this research gap.

[0005] Therefore, how to solve the synchronization error problem while significantly improving the performance of the multi-base station cooperative positioning system is a problem that needs to be solved by those skilled in the art. SUMMARY

[0006] Therefore, the present application provides a multi-base station cooperative robust positioning method and system based on deep learning, which has high-precision positioning capability and robustness to complex channel environments.

[0007] To achieve the above purpose, the present application adopts the following technical solutions: A deep learning-based multi-base station cooperative robust positioning method comprises: Collecting the change channel observation data of the distributed base stations to form a multi-base station multi-time slot joint channel matrix; Performing shift processing on the time slot data of the multi-base station multi-time slot joint channel matrix to simulate the synchronization error of the real scene and obtain a multi-base station joint channel matrix under the synchronization error condition; Performing synchronization error calibration processing on the multi-base station joint channel matrix under the synchronization error condition to obtain a calibrated multi-base station joint channel matrix; Using the calibrated multi-base station joint channel matrix and the built positioning network to perform user coordinate positioning.

[0008] Preferably, the multi-base station joint channel matrix under the synchronization error condition is obtained by applying a cyclic shift operation to the time dimension of the channel state information data in the multi-base station multi-time slot joint channel matrix, wherein there are b base stations in the system, each base station receives channel state information data in t time slots, and the specific operation is represented as: ; Wherein, is a channel state matrix without synchronization error, is a channel state matrix with synchronization error, n is the number of samples, b is the number of base stations, a is the number of antennas, and f is the number of subcarriers, represents the time slot offset of base station b in the nth sample, after applying the cyclic shift operation, the channel state information data is input into the synchronization error calibration network, the number of shifts of the prediction data is predicted and reverse shifting is performed to realize synchronization error calibration.

[0009] Preferably, the synchronization error calibration network comprises: A spatial processing module adopts a three-layer two-dimensional convolution network to extract the subcarrier and antenna dimension spatial features of the channel state information data after the cyclic shift operation, then flattens, obtains a one-dimensional feature vector through two one-dimensional convolution layers, inputs into a first time sequence processing module, and the activation function is LeakyReLU; The first time sequence processing module is based on the MLA mechanism to build a Transformer encoder, compresses the key-value cache through a low-rank latent vector, and captures the time sequence dependency relationship; The classification output layer predicts the time slot offset of each base station through a fully connected layer and a Softmax function; The spatial processing module, the first time sequence processing module and the classification output layer are connected in sequence.

[0010] Preferably, the spatial processing module comprises a first three-layer 2D dilated convolution network, the number of channels is expanded from the initial 2 to 4, 8, 16, and the activation function is LeakyReLU; the first time processing module is a one-dimensional convolution and Transformer encoder dual structure, the features output by the spatial processing module are first input into the first one-dimensional convolution layer, activated by ReLU, then input into the second one-dimensional convolution layer and activated by ReLU again, and then the feature vectors enter the TransformerEncoder module: the core MLA layer directly receives the time sequence of the features; the output of the MLA layer is connected in residual connection with the original input through the Dropout layer, and then input into the LayerNorm layer for normalization; the normalized data is input into the feedforward network; the output of the feedforward network is also connected in residual connection through the Dropout layer and input into the second LayerNorm layer, and finally the time sequence features are compressed into a fixed length vector through the adaptive average pooling layer.

[0011] Preferably, the positioning network comprises a second three-layer 2D dilated convolution network, a second time processing module, and an output module, which are connected in sequence.

[0012] Preferably, the three-layer 2D convolution network extracts spatial features of the channel state information data step by step, and then flattens the spatial features; the second time processing module comprises a one-dimensional convolution network and an xLSTM model, which receives the flattened spatial features through the one-dimensional convolution network and inputs the xLSTM model; the output module receives the output results of the xLSTM model through the time pooling layer, merges and fuses the output results of the time pooling layer, passes through three fully connected layers, and directly outputs continuous coordinate values through the two neuron linear fully connected layer of the last fully connected layer.

[0013] Preferably, the loss function of the synchronization error calibration network is a cross-entropy loss function, which calculates the prediction error of each base station and takes the average value; the loss function of the positioning network is a mean square error loss function, which directly optimizes the coordinate prediction accuracy.

[0014] A multi-base station cooperative robust positioning system based on deep learning, comprising: A matrix establishing module acquires the change channel observation data of the distributed base stations to form a multi-base station multi-time slot joint channel matrix; An error synchronization module performs shift processing on the time slot data of the multi-base station multi-time slot joint channel matrix, simulates the synchronization error of the real scene, and obtains the multi-base station joint channel matrix under the synchronization error condition; A calibration module performs synchronization error calibration processing on the multi-base station joint channel matrix under the synchronization error condition to obtain the calibrated multi-base station joint channel matrix; A positioning module uses the calibrated multi-base station joint channel matrix and the built positioning network to position the coordinates of a user.

[0015] According to the technical solution, compared with the prior art, the application provides a multi-base station cooperative robust positioning method and system based on deep learning, which significantly improves the positioning performance in a complex environment through data-driven error calibration and space-time feature fusion. Simulation results show that the xLSTM model achieves an average absolute positioning error of 0.0347 meters, and the MLA model improves the training efficiency by 6.2 times while maintaining an error of 0.2626 meters. The method provided by the application breaks through the dependence on high-precision clock synchronization equipment and supports robust positioning in dynamic scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0017] Figure 1 A simulation channel model scene graph is provided for the application; Figure 2 A CSI channel model of a multi-base station receiving end is provided for the application; Figure 3 A simulation scene and channel model simulation parameter setting graph is provided for the application; Figure 4 A model structure graph of a synchronization error calibration network is provided for the application; Figure 5 A main structure graph of a transformer encoder used by the model is provided for the application; Figure 6 A multi-base station cooperative positioning system architecture graph is provided for the application; Figure 7a A training process loss change graph of the synchronization error calibration network is provided for the application; Figure 7b A training process accuracy trend graph of the synchronization error calibration network is provided for the application; Figure 8a A comparison graph of xLSTM model positioning errors before and after synchronization error compensation is provided for the application; Figure 8b A comparison graph of MLA model positioning errors before and after synchronization error compensation is provided for the application; Figure 9 A method step flowchart is provided for the application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work belong to the protection scope of the present application.

[0019] The present application mainly includes two stages: the detection and compensation stage of synchronization error, and the multi-base station cooperative positioning stage after compensation. The method aims to solve the data offset problem caused by the clock synchronization error between multiple base stations on CSI (Channel State Information data), and improve the positioning accuracy of wireless users in indoor or complex scenes. The present application provides a multi-base station cooperative robust positioning method based on deep learning. Firstly, the error model based on the CSI time dimension cyclic shift is established according to the clock synchronization error characteristics between base stations, the time sequence dependence is captured by using MLA network, and the error parameters are adaptively estimated. The simulation results show that the displacement prediction accuracy of the calibration network is close to 99.9%, which significantly improves the reliability of time slot alignment. In the design of the positioning network, the architecture of the spatial convolution module and the time sequence processing module (such as xLSTM, MLA) effectively extracts the space-time features of the channel state information data. The xLSTM model realizes an average absolute positioning error of 0.0347 meters after calibration, which verifies the superiority of the method. In addition, by comparing the performance of different time sequence modules, the advantages of xLSTM in accuracy and the balance of MLA in efficiency are revealed, which provides theoretical support for model selection in actual scenarios. Not only does it provide a new technical path for multi-base station cooperative positioning under the condition of synchronization error, but also verifies the applicability of data-driven methods in complex channel environments, which has important engineering value for high-precision positioning scenarios such as intelligent transportation and indoor navigation.

[0020] The embodiments of the present application disclose a multi-base station cooperative robust positioning method based on deep learning, as shown in Figure 9 , which comprises: Collecting the changing channel observation data of distributed base stations to form a multi-base station multi-time slot joint channel matrix; Performing shift processing on the time slot data of the multi-base station multi-time slot joint channel matrix to simulate the synchronization error of the real scene, and obtaining the multi-base station joint channel matrix under the condition of synchronization error; Performing synchronization error calibration processing on the multi-base station joint channel matrix under the condition of synchronization error to obtain the calibrated multi-base station joint channel matrix; Using the calibrated multi-base station joint channel matrix and the built positioning network to position the user coordinates.

[0021] In one specific embodiment, obtaining the multi-base station joint channel matrix under synchronization error conditions comprises applying a cyclic shift operation to the time dimension of the channel state information data in the multi-base station multi-time slot joint channel matrix, assuming that there are b base stations in the system, and each base station receives channel state information data in t time slots. The specific operation is represented as: ; wherein, is the channel state matrix without synchronization error, is the channel state matrix with synchronization error, n is the number of samples, b is the number of base stations, a is the number of antennas, and f is the number of subcarriers, represents the time slot offset of base station b in the nth sample. After applying the cyclic shift operation, the channel state information data is input into the synchronization error calibration network, and the synchronization error is calibrated by predicting the number of shifts and performing reverse shifting.

[0022] In one specific embodiment, the synchronization error calibration network comprises: a spatial processing module that extracts the subcarrier and antenna dimension spatial features of the channel state information data after the cyclic shift operation using a three-layer two-dimensional convolution network, then flattens, obtains a one-dimensional feature vector after two one-dimensional convolution layers, inputs the first time series processing module, and selects LeakyReLU as the activation function; a first time series processing module, i.e. Figure 4 the Transformer encoder in the above formula, the Transformer encoder is constructed based on the MLA mechanism, the key-value cache is compressed through the low-rank latent vector, the inference memory occupation is reduced, and the time series dependency is captured; a classification output layer that predicts the time slot offset of each base station using a fully connected layer and a Softmax function; The spatial processing module, the first time series processing module, and the classification output layer are connected in sequence.

[0023] Note that in the aforementioned spatial processing module and the first time series processing module, the target is the signal matrix of a single base station. Therefore, after time series processing, the signal matrices of different base stations of the same target are combined, and then passed through the classification output layer. In one specific embodiment, the spatial processing module comprises a first three-layer 2D dilated convolution network, the number of channels is expanded from the initial 2 to 4, 8, 16, and the activation function is LeakyReLU; the first temporal processing module is a dual structure of one-dimensional convolution and Transformer encoder, the features output by the spatial processing module are first input into the first one-dimensional convolution layer, activated by ReLU, then input into the second one-dimensional convolution layer and activated by ReLU again, and then the feature vector enters the TransformerEncoder module: the core MLA layer directly receives the sequence of temporal features; the output of the MLA layer is connected in residual connection with the original input through the Dropout layer, and then input into the LayerNorm layer for normalization; the normalized data enters the feedforward network; the output of the feedforward network is also connected in residual connection through the Dropout layer and input into the second LayerNorm layer, and finally the temporal features are compressed into a fixed length vector through the adaptive average pooling layer.

[0024] In one specific embodiment, the positioning network comprises a second three-layer 2D dilated convolution network, a second temporal processing module, and an output module, which are connected in sequence.

[0025] In one specific embodiment, the three-layer 2D convolution network extracts spatial features of channel state information data step by step, and then flattens the spatial features; the second temporal processing module comprises a one-dimensional convolution network and an xLSTM model, which receives the flattened spatial features through the one-dimensional convolution network and inputs them into the xLSTM model; the xLSTM model adopts an exponential gating and matrix memory structure, and realizes parallel processing through a covariance update rule; the output module receives the results output by the xLSTM model through a temporal pooling layer, merges and fuses the output results of the temporal pooling layer, passes through three fully connected layers, and directly outputs continuous coordinate values through a two-neuron linear fully connected layer in the last fully connected layer.

[0026] In one specific embodiment, the loss function of the synchronization error calibration network is a cross-entropy loss function, which calculates the prediction error of each base station and takes the average value; the loss function of the positioning network is a mean square error loss function, which directly optimizes the coordinate prediction accuracy.

[0027] A multi-base station cooperative robust positioning system based on deep learning, comprising: A matrix establishment module acquires the change channel observation data of the distributed base stations to form a multi-base station multi-time slot joint channel matrix; An error synchronization module performs shift processing on the time slot data of the multi-base station multi-time slot joint channel matrix to simulate the synchronization error of the real scene and obtain the multi-base station joint channel matrix under the synchronization error condition. The calibration module performs synchronization error calibration processing on the multi-base station joint channel matrix under the synchronization error condition to obtain a calibrated multi-base station joint channel matrix. The positioning module uses the calibrated multi-base station joint channel matrix and the built positioning network to position the user coordinates.

[0028] The simulation verification condition of the application is: setting the working frequency band as 3.5 GHz, the bandwidth as 40 MHz, and the subcarrier number as 64, generating a CSI data set containing random time slot offset; using a ray tracing algorithm to simulate multipath propagation, constructing a time-varying channel impulse response and converting it into frequency domain CSI data; comparing and analyzing the positioning errors before and after calibration to verify the effectiveness of the synchronization error compensation. At the same time, for dynamic scenarios, the resource allocation is optimized in combination with 6G ISAC (Integrated Sensing and Access) technology to improve the environmental adaptability of the positioning system.

[0029] The application builds a real urban street ray tracing scene in the Sionna simulation platform, as shown in Figure 1 The CSI data of the base station and the user equipment are obtained by ray tracing, and the data set covers multiple synchronization error levels and user motion states. By comparing the error performance of the CSI data in the positioning network before and after calibration, the effectiveness of the synchronization error calibration method of the application is significantly verified.

[0030] As shown in Figure 2 The application establishes a CSI channel model of the multi-base station receiving end in the system model and constructs a five-dimensional CSI tensor with a time dimension. To simulate the synchronization error between different base stations, the CSI tensor is quantitatively shifted in the time dimension to simulate the receiving timing difference between base stations due to clock offset in the actual environment, thereby constructing a data set with offset labels for subsequent error detection network training.

[0031] In the training process of the synchronization error calibration network, the time slot offset is set as discrete values of ±1 to ±4 to simulate microsecond-level clock deviation, the classification head parameter quantity linearly increases with the offset range, and the total memory occupation is 294.55~294.57 MB.

[0032] The main parameter configuration of the application is shown in Figure 3 The core parameter configuration of the simulation system is mainly set based on two elements of actual demand and algorithm verification target of the wireless communication scene. The system working frequency band is set as 3.5 GHz to cover the typical application scenario of medium frequency band wireless communication. The total bandwidth of the system is 40 MHz, which is divided into 64 subcarriers by orthogonal frequency division multiplexing technology, and the subcarrier interval is calculated by uniformly allocating the total bandwidth as 625 kHz.

[0033] To simulate the multipath effect in complex propagation environment, the maximum number of reflections is set to 5, covering the physical phenomena such as reflection and scattering experienced by the signal during propagation.

[0034] The antenna configuration of the base station and the user equipment is an important part of the simulation scenario. The base station receiver uses a 2-row x 4-column planar array antenna with a vertical and horizontal spacing of 0.5 wavelengths and a vertical polarization to enhance the spatial diversity capability. The user equipment uses a single antenna transmission configuration to simulate the simplified hardware design of a typical mobile terminal.

[0035] To accurately capture the Doppler effect caused by the movement of the user equipment, the simulation sets the sampling frequency to 625 kHz (matching the subcarrier allocation bandwidth calculated earlier) and the time step to 14. Through the reciprocal relationship between period and frequency, it can be calculated that the time difference between adjacent time slots is 1.6 microseconds.

[0036] The embodiment of the present application discloses a synchronization error calibration network model, as shown in Figure 4 The transformer encoder structure used is as shown in Figure 5 .

[0037] The neural network model is composed of a spatial processing module, a first time series processing module and an output layer, and each module works cooperatively through a specific connection method. The spatial processing module adopts a three-dimensional convolution structure. First, the input data is processed by a first two-dimensional convolution layer (input channel 2, output channel 4, 3x3 kernel), then it is nonlinearly transformed by a LeakyReLU activation function (negative slope 0.1); its output is directly connected to a second two-dimensional convolution layer (input channel 4, output channel 8, 3x3 kernel), which is also connected to a LeakyReLU activation function; then it enters a third two-dimensional convolution layer (input channel 8, output channel 16, 3x3 kernel) and is again passed through a LeakyReLU activation function. Finally, the output of this module is compressed into a one-dimensional feature vector by a flattening layer (Flatten).

[0038] The first time sequence processing module comprises a one-dimensional convolution and a dual structure of a Transformer encoder. The features output by the spatial processing module are first input into a first one-dimensional convolution layer (the input channel is the automatically calculated spatial feature dimension, the output channel is 512, and the kernel size is 3), activated by ReLU, and then input into a second one-dimensional convolution layer (the input channel is 512, the output channel is 64, and the kernel size is 3) and activated by ReLU again. Subsequently, the feature vector enters the TransformerEncoder module: the core MLA layer directly receives the time sequence feature sequence; the MLA output is connected in residual connection with the original input through a Dropout layer (the dropout rate is 0.1), and then input into a LayerNorm layer for normalization; the normalized data enters a feedforward network (FFN), which is sequentially connected by a linear layer (the input / output dimension is 64 / 128), a ReLU activation function, a Dropout layer (the dropout rate is 0.1), and a linear layer (128 / 64); the FFN output is also connected in residual connection through a Dropout layer and input into a second LayerNorm layer. Finally, the time sequence feature is compressed into a fixed length vector through an adaptive average pooling layer (AdaptiveAvgPool1d).

[0039] The output layer is for the discrete classification task of the time slot offset of six base stations. First, the feature vectors of the six base stations are spliced and input into a classifier sequence: a first linear layer (the input dimension is 64x6, and the output is 128) is connected with a ReLU activation function, and a second linear layer (128 dimensions to 6xnum_classes output). Finally, the output tensor is reshaped into a three-dimensional structure (batchx6 sitesxclass number), to complete the end-to-end mapping from the original space-time input to the site-level prediction. In the whole processing process, the spatial convolution extracts local features, the time sequence convolution captures short-range dependencies, the Transformer encoder models long-range time sequence relationships, and finally the multi-site joint prediction is realized through a fully connected layer. In the simulation, the time slot shift range covers ±1 to ±4, corresponding to the microsecond-level synchronization error, and the precise error compensation is realized through classification prediction.

[0040] The embodiment of the application discloses a positioning network model, such as Figure 6As shown, the model also extracts the spatial features of the CSI data step by step through a second three-layer 2D dilated convolutional network. The spatial correlation of the signal in the subcarrier and path dimensions is captured through multi-level convolutional operations. After flattening the spatial features, the time series is processed through a one-dimensional convolutional network to reduce the computational load and realize feature integration. The second time series processing module directly accesses the xLSTM model without the need to build a transformer encoder to assist. The multi-base station cooperative positioning network needs to predict the two-dimensional spatial coordinates of the user equipment as a regression task. Therefore, the output layer adopts a two-neuron linear fully connected layer to directly output continuous coordinate values. This design avoids unnecessary interference of the nonlinear activation function on the regression value and ensures the numerical continuity of the coordinate prediction.

[0041] Figure 7a The loss changes in the training processes of the four models are shown in the figure, Figure 7b The accuracy trend in the training processes of the four models is shown in the figure. The loss functions of the models all show a downward trend with the increase of the iteration number, and all tend to be stable after 100 rounds of training, indicating good convergence of the models.

[0042] Figure 8a With Figure 8b The performance of different network structures under various synchronization error conditions is shown. The experimental results show that: 1. The calibration network can accurately identify and correct the time offset between base stations, and the restored CSI data significantly reduces the positioning error. Especially in the high error scenario (±4 time slot shift), the positioning accuracy after calibration can basically recover to the error-free level; 2. The positioning network can achieve high-precision (centimeter-level) positioning in the current scenario, and has certain tolerance to synchronization error.

[0043] In summary, the multi-base station cooperative positioning method and system under synchronization error conditions provided by the application comprehensively use deep learning technologies such as cyclic shift modeling, multi-head latent attention mechanism (MLA) and extended long short-term memory network (xLSTM), realize accurate compensation of synchronization error between multiple base stations and significant improvement of user positioning accuracy. Through the extraction and regression optimization of the space-time features of the CSI data, the application shows good robustness and engineering adaptability in complex multipath environments and dynamic scenarios.

[0044] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0045] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-base station cooperative robust localization method based on deep learning, characterized in that, include: Collect change channel observation data from distributed base stations to form a multi-base station, multi-timeslot joint channel matrix; The time slot data of the multi-base station multi-time slot joint channel matrix is ​​shifted to simulate the synchronization error in a real-world scenario, and the multi-base station joint channel matrix under the synchronization error condition is obtained. The multi-base station joint channel matrix under the aforementioned synchronization error condition is subjected to synchronization error calibration processing to obtain the calibrated multi-base station joint channel matrix. User coordinates are located using the calibrated multi-base station joint channel matrix and the established positioning network.

2. The multi-base station cooperative robust localization method based on deep learning according to claim 1, characterized in that, The acquisition of the multi-base station joint channel matrix under synchronization error conditions includes applying a cyclic shift operation to the time dimension of the channel state information data in the multi-base station multi-timeslot joint channel matrix. Assuming there are b base stations in the system, and each base station receives channel state information data within t time slots, the specific operation is as follows: ； in, It is a channel state matrix without synchronization error. This is the channel state matrix with synchronization error, where n is the number of samples, b is the number of base stations, a is the number of antennas, and f is the number of subcarriers. This represents the time slot offset of base station b in the nth sample. After applying a cyclic shift operation, the channel state information data is input into the synchronization error calibration network. Synchronization error calibration is achieved by predicting the number of shifts in the data and performing a reverse shift.

3. The multi-base station cooperative robust localization method based on deep learning according to claim 2, characterized in that, The synchronization error calibration network includes: The spatial processing module uses a three-layer two-dimensional convolutional network to extract the subcarrier and antenna spatial features of the channel state information data after cyclic shifting. Then, it is flattened and passed through two layers of one-dimensional convolution to obtain a one-dimensional feature vector, which is then input into the first time-series processing module. The activation function is LeakyReLU. The first time-series processing module constructs a Transformer encoder based on the MLA mechanism and captures time-series dependencies by compressing key-value caches using low-rank latent vectors. The classification output layer mainly uses a fully connected layer and a Softmax function to predict the time slot offset of each base station. The spatial processing module, the first temporal processing module, and the classification output layer are connected in sequence.

4. The multi-base station cooperative robust localization method based on deep learning according to claim 3, characterized in that, The spatial processing module includes a first three-layer 2D dilated convolutional network, with the number of channels expanded from the initial 2 to 4, 8, and 16, and LeakyReLU activation function selected. The first temporal processing module is a dual structure of one-dimensional convolution and Transformer encoder. The features output by the spatial processing module are first input into the first one-dimensional convolutional layer, activated by ReLU, and then input into the second one-dimensional convolutional layer and activated by ReLU again. Subsequently, the feature vector enters the TransformerEncoder module: its core MLA layer directly receives the temporal feature sequence; the output of the MLA layer is residually connected to the original input through the Dropout layer, and then input into the LayerNorm layer for normalization; the normalized data enters the feedforward network; the output of the feedforward network is also residually connected through the Dropout layer and input into the second LayerNorm layer. Finally, the temporal features are compressed into a fixed-length vector through an adaptive average pooling layer.

5. The multi-base station cooperative robust localization method based on deep learning according to claim 1, characterized in that, The positioning network includes a second and third layer 2D dilated convolutional network, a second temporal processing module, and an output module, which are connected in sequence.

6. The multi-base station cooperative robust localization method based on deep learning according to claim 2, characterized in that, The loss function of the synchronization error calibration network is the cross-entropy loss function, which calculates the prediction error of each base station and takes the average value; the loss function of the positioning network is the mean square error loss function, which directly optimizes the coordinate prediction accuracy.

7. A multi-base station cooperative robust localization system based on deep learning, employing the multi-base station cooperative robust localization method based on deep learning as described in any one of claims 1-6, characterized in that, include: The matrix building module collects changing channel observation data from distributed base stations and forms a multi-base station, multi-timeslot joint channel matrix. The error synchronization module performs shift processing on the time slot data of the multi-base station multi-time slot joint channel matrix to simulate the synchronization error in the real scenario and obtain the multi-base station joint channel matrix under the synchronization error condition. The calibration module performs synchronization error calibration on the joint channel matrix of multiple base stations under synchronization error conditions to obtain the calibrated joint channel matrix of multiple base stations. The positioning module uses the calibrated multi-base station joint channel matrix and the established positioning network to locate the user's coordinates.

Images