Network large-scale MIMO multi-base-station joint positioning method

Abstract

The invention discloses a network large-scale MIMO (Multiple Input Multiple Output) multi-base station joint positioning method, which comprises the following steps that: each base station extracts a low-dimensional angle delay domain sparse matrix (ADSM) as a position fingerprint by utilizing statistical channel information, trains a first network model by taking the ADSM fingerprint and a corresponding position coordinate as a data set, preliminarily extracts features, and obtains a second network model; the feature represents a preliminary prediction result of each base station on the user position. And fusing the features extracted by all the single base stations to construct a joint fingerprint, training a second network model by using joint fingerprint information and corresponding position coordinates as a data set, constructing a joint fingerprint for a to-be-positioned user according to the position fingerprint extracted by the single base station, and estimating the user position by using a multi-base-station joint positioning module. According to the multi-base-station combined positioning method provided by the invention, the storage dimension of fingerprint information and the network overhead of position estimation can be effectively reduced, and the positioning precision is effectively improved through a multi-base-station cooperation mode.

Description

Technical Field

[0001] The present invention relates to a positioning technology, in particular to a network large-scale MIMO multi-base station joint positioning method, and belongs to the field of mobile communication technology. Background Art

[0002] Obtaining the location information of user terminals (UT) is very important for many smart city and IoT applications, such as traffic monitoring, asset tracking, autonomous driving, emergency rescue, etc. However, the obstruction of the line of sight path seriously degrades the performance of global positioning systems in complex urban canyons or indoor environments. In recent years, positioning methods based on location fingerprint information are attracting increasing research enthusiasm due to their wide range of application scenarios. Massive multiple-input multiple-output (MIMO) and orthogonal frequency division multiplexing (OFDM) transmission technologies are key components of the fifth generation (5G) wireless cellular systems, and location fingerprint information positioning methods can tap the potential of multipath propagation to improve the positioning accuracy of massive MIMO-OFDM systems.

[0003] So far, there have been some positioning methods based on location fingerprints. For example, the angle delay domain channel power matrix (ADCPM) is proposed as the location fingerprint. The two-level fingerprint clustering method is used to compare the user's location fingerprint with the most similar fingerprint in the database, and then the relevant location is returned; the angle delay domain channel amplitude matrix (ADCAM) is used as the location fingerprint, and the deep convolutional neural network (DCNN) is applied to improve the fingerprint matching of location prediction; ADCPM and its variants are used as location fingerprints to achieve user positioning with the assistance of machine learning. However, these positioning methods based on location fingerprints are relative to a single base station (BS). Due to factors such as scatterer occlusion and the distance of the user to be located, the positioning under a single base station often has more uncertainties and large positioning errors. 5G is expected to achieve high spatial density of base stations in urban scenarios, supporting high-speed communication and high-precision positioning services. Multiple base stations can cooperate with each other to make up for the shortcomings of the large positioning error of a single base station and achieve more accurate positioning. Compared with single base station positioning, multi-base station positioning has great advantages in positioning accuracy, positioning reliability, positioning coverage, and positioning speed. In addition, existing more mature fingerprint-based positioning methods usually use statistical channel response matrices (such as ADCPM) as location fingerprints, and their dimensions are usually large, which will lead to huge storage overhead and time complexity. Therefore, it is necessary to reduce the dimension of location fingerprints.

[0004] Currently, the location fingerprint-based positioning method usually adopts machine learning because the location fingerprint has complex non-linear relationships. Machine learning algorithms can well capture these non-linear features, discover hidden patterns and rules from seemingly chaotic data, and then accurately predict the location of the terminal to be located. The generation of the offline reference point fingerprint database and the online UT location prediction can be modeled as a training and prediction process, similar to image recognition. Regarding the UT fingerprint as the picture to be recognized, training the fingerprint dataset to learn the features between the fingerprint and the location, and transforming the positioning problem into an image regression problem in computer vision (CV) to achieve the prediction of the location of the terminal to be located. In machine learning algorithms, traditional neural networks may have defects such as distorted predicted locations and low prediction location accuracy in the task of predicting locations. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide a network large-scale MIMO multi-base station joint positioning method to reduce the storage overhead and improve the positioning accuracy.

[0006] Technical Solution: To achieve the above object, the present invention adopts the following technical solution:

[0007] The network large-scale MIMO multi-base station joint positioning method includes the following steps:

[0008] Each base station extracts a low-dimensional angular delay domain sparse matrix (ADSM) as the location fingerprint according to the statistical channel information, and uses the ADSM fingerprint and the corresponding location coordinates as the dataset to train the first network model to initially extract features;

[0009] Fuse the features extracted by all single base stations to construct a joint fingerprint, and use the joint fingerprint and the corresponding location coordinates as the dataset to train the second network model. For the user to be located, construct a joint fingerprint based on the features extracted by all single base stations through the first network model, and use the trained second network model to estimate the user location.

[0010] Further, the angular delay domain sparse matrix (ADSM) is obtained by dimension reduction transformation of the angular delay domain channel energy matrix Ω, where N s represents the number of peak elements in the original Ω matrix; the three columns of Ψ respectively represent the row index, column index of the peak elements in the Ω matrix, and the value of the element at the corresponding row and column index positions.

[0011] Furthermore, the first network model is a Feature Separation Generative Adversarial Network (FSGAN), which includes a generator G and a discriminator D. The generator includes a mapping network and a synthesis network. FSGAN combines GAN and supervised learning by adding labels and uses conditional information to guide the generation process. In the position prediction task, the fingerprint condition Ψ related to the position Ξ is used as the input, and by learning the distribution p data (Ξ|Ψ) of the real data and generating similar but not exactly the same data p g (Ξ|Ψ).

[0012] Furthermore, the FSGAN adopts progressive training to capture the potential relationship between position data and related fingerprints. The synthesis network contains multiple sub-networks, and the sample dimensions of each sub-network are different, increasing step by step to refine the learned features.

[0013] Furthermore, during the training process, the generator and discriminator are first trained using low-dimensional noise samples, and then each time it ascends to a higher-dimensional sub-network. During the dimension increase process, a mixing coefficient is introduced for smooth transition.

[0014] Furthermore, the mapping network consists of multiple fully connected layers and is used to decouple the input fingerprint into an intermediate hidden variable w. The synthesis network is composed of multiple sub-networks with progressive training. The input of the first-layer sub-network is a constant, and the input of the remaining sub-networks is the output of the previous-layer sub-network. The first-layer sub-network includes a convolutional layer, and the remaining sub-networks include two convolutional layers and an upsampling layer. Transformed random noise and an affine transformation of w are added to each layer of the sub-network. Each dimension level of the synthesis network is affected by the affine transformation of w, and the influence method uses batch normalization.

[0015] Furthermore, the second network model is defined as a joint FSGAN model, and its structure is similar to that of the FSGAN of the first network model.

[0016] Furthermore, the joint fingerprint information of the reference point and its position coordinates are used as training data to train the second network model, and the trained network structure is stored in the database. During multi-base station joint positioning, each base station extracts ADSM position fingerprint information from the channel state information estimated from the uplink channel, and uses its own trained first network model to extract the position features of the terminal to be located. The position features extracted by each base station for the terminal to be located are fused into a joint fingerprint of the terminal to be located and input into the trained second network model to output the final predicted position of the terminal to be located.

[0017] A computer system includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the steps of the network large-scale MIMO multi-base station joint positioning method are implemented.

[0018] A computer program product includes a computer program. When the computer program is executed by the processor, the steps of the network large-scale MIMO multi-base station joint positioning method are implemented.

[0019] Beneficial effects: Compared with the prior art, a network large-scale MIMO multi-base station joint positioning method disclosed by the present invention has the following advantages:

[0020] 1. Small storage overhead: The ADSM location fingerprint disclosed by the present invention is the result of dimensionality reduction of the angle-delay domain channel energy matrix (ADCPM) commonly used in existing positioning methods through a compression algorithm. In addition, the dimension of the joint fingerprint is low, so the storage overhead of fingerprint information can be effectively reduced.

[0021] 2. Low time complexity: The ADSM location fingerprint and the joint fingerprint disclosed by the present invention have low dimensions, which can effectively reduce the network overhead of training the model, so the time complexity is reduced.

[0022] 3. High positioning accuracy: The multi-base station joint positioning method proposed by the present invention combines the features extracted by multiple single base stations, which can effectively improve the deficiency of low positioning accuracy of single base stations. In addition, the further improved FSGAN model of the present invention can fully learn the sample features by adding conditional information constraints, adopting progressive training, and separating features using the mapping network and the synthesis network. The generalization performance of the model is good, and the positioning accuracy can be improved. Description of the Drawings

[0023] Figure 1 It is an example diagram of the angle-delay domain channel intensity matrix in the embodiment of the present invention, N = 128, N g = 144;

[0024] Figure 2 It is a block diagram of the positioning method in the embodiment of the present invention;

[0025] Figure 3 It is a structure diagram of FSGAN in the embodiment of the present invention;

[0026] Figure 4 It is a comparison diagram of the positioning accuracy of different fingerprints;

[0027] Figure 5 It is a comparison diagram of the positioning accuracy between the positioning method in the embodiment of the present invention and the existing method. Detailed Embodiments

[0028] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0029] In order to reduce the storage overhead and improve the positioning accuracy, an embodiment of the present invention discloses a network large-scale MIMO multi-base station joint positioning method, including single-base station feature extraction and multi-base station joint positioning. In single-base station feature extraction, each base station extracts a low-dimensional angular-delay domain sparse matrix (ADSM) as a location fingerprint according to the statistical channel information, and uses the ADSM fingerprint and the corresponding location coordinates as a data set to train a first network model to initially extract features, which represent the initial prediction results of each base station for the user location. In multi-base station joint positioning, the features extracted by all single-base stations are fused to construct a joint fingerprint, and the joint fingerprint information and the corresponding location coordinates are used as a data set to train a second network model. For the user to be located, a joint location fingerprint is constructed based on the features extracted by all single-base stations, and the trained model is used to estimate the user location. The detailed implementation process of the network large-scale MIMO multi-base station joint positioning method will be described below in conjunction with a specific system model.

[0030] Considering a typical large-scale MIMO-OFDM system, it is assumed that single-antenna user terminals are randomly distributed in the area to be located, and there are multiple base stations in the area. Each base station is configured with a uniform linear array (ULA) composed of N antennas. For simplicity and without loss of generality, it is assumed that the antenna spacing d is half of the carrier wavelength λ, that is The number of subcarriers of the OFDM system is N c , and the total bandwidth is B. It is assumed that L scatterers are randomly distributed in the area to be located.

[0031] Considering the physical propagation model of the channel, the steering vector corresponding to the angle of arrival (AoA) θ is defined as

[0032] a(θ) = (1 e -jπcos(θ) …e -jπ(N-1)cos(θ) ) T

[0033] It is assumed that the signal between the user and the kth base station is transmitted through L k paths. Let θ k，l be defined as the AoA of the lth path between the user and the kth base station, and τ k，l be defined as the time of arrival (ToA, delay of arrival) of the lth path between the user and the kth base station. The channel frequency response (CFR) at the nth subcarrier of the kth base station can be expressed as

[0034]

[0035] where is the fading coefficient of the l-th path between the user and the k-th base station, f n represents the frequency of the n-th subcarrier. The overall channel matrix between the user and the k-th base station is denoted as h k，n stacked as:

[0036]

[0037] For the positioning method based on location fingerprint information, the location fingerprint information extracted from the channel state information needs to satisfy the following constraints:

[0038] (1) The location fingerprint information needs to be uniquely determined by the scatterer environment around the mobile terminal.

[0039] (2) The location fingerprint information needs to maintain a wide-sense stationary process from the offline stage to the online stage.

[0040] (3) The location fingerprint information needs to have sufficient distinguishability between different geographical locations.

[0041] Based on the above two constraints, the CFR at the k-th base station side is mapped to an angular-delay domain channel response matrix (ADCRM) with a sparse structure through DFT

[0042]

[0043] where denotes the matrix composed of the first N c columns of the N dimensional unitary DFT matrix g N g satisfies N g = T g / T s where T g represents the cyclic prefix length and T s represents the sampling interval. Each of its elements The left-side algorithm multiplier V H and the right-side algorithm multiplier map the spatial frequency domain CFR to the angular domain and the delay domain respectively. Therefore, the (i, j)-th element of the ADCRM represents the complex channel gain corresponding to the i-th AoA direction at the j-th ToA sampling point.

[0044] To comprehensively describe the wide-sense stationary characteristics such as the energy, AoA, and ToA of each path corresponding to the scatterer environment where the user is located, the angular-delay domain channel energy matrix (ADCPM) at the k-th base station side is defined as

[0045]

[0046] wherein, ⊙ represents the Hadamard product,

[0047] In a real - world scenario, the channel energy σ corresponding to different paths k，l will have a large gap. As a result, when calculating the location fingerprint information, several main paths with larger channel energy will occupy a great weight, while paths with weaker channel energy will hardly affect the discrimination of the location fingerprint information. This corresponds to k in the Ω Figure 1 matrix, the peaks are concentrated in a certain part of the region, and most of the other values are almost 0, as k shown. This leads to a very low utilization rate of the Ω k matrix because a large number of values close to 0 at different positions have no practical significance, and the discrimination between different geographical locations is reflected at the peak positions. Therefore, in order to improve the utilization rate of ADCPM, the embodiments of the present invention transform the Ω matrix into a sparse matrix s and define it as the angular - delay - domain sparse matrix (ADSM), where N k represents the number of peak elements in the original Ω k matrix. Here, the range of the peak is defined as exceeding ∈% of the maximum value in the Ω k matrix (∈ represents a threshold value). The first column of Ψ k represents the rows (delay - domain sampling points) of the peak elements in the Ω k matrix, the second column represents the columns (angular - domain sampling points) of the peak elements in the Ω k matrix, and the third column represents the values (response amplitudes) of the peak elements in the Ω

[0048] Taking the ADCPM matrix Figure 1 in as an example, its dimension is as high as 128×144 = 18432, and the number of peak elements in the matrix is 292 (obtained by taking ∈ = 1). After being transformed into a sparse matrix , the dimension is 292×3 = 876. It can be seen that the dimension of Ψ k is only about 5% of that of Ω k , and the dimension - reduction effect is significant. Using ADSM as the new location fingerprint information for positioning, compared with the ADCPM location fingerprint information, the ADSM location fingerprint information retains the peak elements in the original matrix, that is, the discrimination between different geographical locations, and at the same time greatly reduces the dimension, that is, greatly reduces the storage overhead, providing favorable conditions for subsequent machine - learning training.

[0049] The positioning accuracy of a single base station is greatly affected by factors such as scatterer occlusion and the distance of the user, and the positioning error is obvious. In an embodiment of the present invention, a combined fingerprint is constructed, which is a fusion of features extracted from all single base stations. Taking three base stations as an example, when each base station locates the user independently, there will be respective positioning errors. However, if the features extracted from these three base stations (denoted as Ξ1, Ξ2, and Ξ3 for example) are combined in a certain relationship to obtain the predicted position of the user under multi-base-station cooperation

[0050]

[0051] Ξ will be closer to the true position of the user. This is because due to factors such as fading and scatterer occlusion, the positioning accuracy of the base station for a user at a closer distance is higher, and the positioning accuracy for a user at a farther distance is relatively lower. Correspondingly, after each base station has achieved its own feature extraction, the base station closer to the true position of the user, the corresponding feature Ξ k , k ∈ {1, 2, 3} has a greater influence on the final predicted position Ξ. Make Ξ k , k ∈ {1, 2, 3} have a mutually restrictive relationship, and a higher-precision positioning of the user is achieved during the mutual restriction process. The multi-base-station joint positioning problem in an embodiment of the present invention can be transformed into the following two sub-problems:

[0052] ·ADSM location fingerprint extraction and single-base-station feature extraction.

[0053] ·Construction of the combined fingerprint and realization of multi-base-station joint positioning.

[0054] Figure 2 This is the overall process of the positioning method in an embodiment of the present invention, and the whole process includes two stages: single-base-station feature extraction and multi-base-station joint positioning.

[0055] Regarding the single-base-station feature extraction described in the first sub-problem: Reference points are uniformly set in the area to be located. The unmanned vehicle samples the mobile terminal offline when it passes through each reference point, and collects the position fingerprint information and position coordinate information of each base station at each reference point. Each base station uses the reference point position as the label of the reference point, and uses the position fingerprint information of the reference point as the training data to train the FSGAN. The obtained FSGAN node weights and biases and the position fingerprint information dataset are stored in the base station side database. Since both the training data collection and the FSGAN training are carried out in the offline stage, the computational complexity and latency of the position prediction in the online stage are greatly reduced. In addition, each base station uses its own trained FSGAN to extract features from the reference points. Taking three base stations as an example in an embodiment of the present invention, the features extracted by each base station from the reference points are denoted as Ξ1, Ξ2, and Ξ3, and a combined fingerprint is constructed (The features extracted here are two-dimensional position coordinates for exemplary illustration) and stored in the database.

[0056] For the multi-base station joint positioning described in the second sub-question, the joint fingerprint information of the reference point and its position coordinates are used as training data to train the joint FSGAN model. The joint FSGAN model corresponds to the information. The network structure of the trained joint FSGAN is stored in the database. During multi-base station joint positioning, each base station extracts the ADSM position fingerprint information from the channel state information estimated from the uplink channel, and uses its own trained FSGAN to extract the position features of the terminal to be located. The position features extracted by each base station for the terminal to be located are fused into the joint fingerprint of the terminal to be located and input into the trained joint FSGAN model, and the final predicted position of the terminal to be located is output

[0057] The embodiment of the present invention constructs an FSGAN model to estimate the user position. Figure 3 is the designed FSGAN model diagram. The structure of the joint FSGAN is similar to that of the FSGAN.

[0058] FSGAN is a deep neural network architecture, mainly including two parts: a generator G and a discriminator D. The generator is mainly used to learn the real sample distribution so that the data generated by itself is more real to deceive the discriminator. It includes two parts: a mapping network and a synthesis network; the discriminator needs to judge the authenticity of the received data. During the training process, the discriminator and the generator update their parameters alternately. First, fix the parameters of the generator and update the parameters of the discriminator to minimize the loss L D of the discriminator; then fix the parameters of the discriminator and update the parameters of the generator to minimize the loss L G of the generator. Through continuous alternating training, the capabilities of the generator and the discriminator are gradually improved, and finally a relatively balanced state is reached.

[0059] By adding labels, FSGAN combines GAN and supervised learning, and uses conditional information to guide the generation process, making GAN more controllable. In position prediction, the fingerprint condition Ψ related to the position Ξ is used as the input. Through adversarial training, the generator is continuously optimized during the continuous game between the generator and the discriminator. Finally, the generated data distribution p g (Ξ|Ψ) gradually approaches the real data distribution p data (Ξ|Ψ). By learning the distribution p data (Ξ|Ψ) of the real data and generating data p g(Ξ|Ψ), FSGAN can improve the generalization ability of the model, enabling it to generate outputs that conform to a specific distribution and make relatively accurate location predictions even when faced with unknown fingerprint conditions. FSGAN adopts progressive training to more accurately capture the potential relationship between location data and relevant fingerprints. The synthesis network consists of multiple sub-networks, each with a different sample dimension, which gradually increases level by level to refine the learned features; during the training process, first, the generator and discriminator are trained using low-dimensional noise samples (such as 4×4×number of channels), and then each time it ascends to a sub-network with a higher dimension; for example, when ascending from the 4×4 dimension to the 8×8 dimension, a mixing coefficient α is introduced for smooth transition. When α = 0, the generator mainly relies on the previously trained 4×4 network part, and through upsampling, it obtains 8×8 feature samples, and then a new convolutional layer is added to extract features from the 8×8 samples; as the training progresses, α gradually increases, and the contribution of the newly added 8×8 convolutional layer gradually increases. When α = 1, the samples are completely generated by the newly added 8×8 convolutional layer; this smooth transition process can be achieved through a weighted summation operation, that is, G output = α × G 8×8 +(1 - α) × Upsamole(G 4×4 ), where G output is the finally output 8×8 sample, G 8×8 is the sample generated by the newly added 8×8 convolutional layer, G 4×4 is the sample generated by the previous 4×4 network. Upsample is the upsampling operation. This technique first forms the basis for generating data by learning the basic features that appear in low-dimensional samples, and then, as the dimension increases, it learns more and more details as the training progresses. Training low-dimensional samples is not only easier and faster but also helps train higher-level samples (preventing falling into sub-optimal solutions). Therefore, doing so not only enables greater stability in the early stage of training but also brings a faster training speed.

[0060] The goal of the generator G is to deceive the discriminator D into believing that the generated data G(z, Ψ) is real, where z is random noise and Ψ is fingerprint information. During the training process, the generator adjusts its parameters so that the value of D(G(z, Ψ), Ψ) (the probability that the discriminator determines the generated data to be true) is as close to 1 as possible, that is, making it difficult for the discriminator to distinguish between the generated data and the real data. The corresponding loss function is The generator will update its parameters according to the gradient information of the loss function, optimize in the direction of reducing the loss function, and continuously adjust the generated data distribution to make it closer to the real data distribution.

[0061] Mapping Network: It consists of 4 fully connected layers. What it does is to decouple the input data (fingerprint) into the intermediate hidden variable w. The features of the input data are correlated with each other and have a relatively high coupling. w can relatively well represent the results after decoupling these features, enabling the model to better learn the relationships between features and improve the learning efficiency. Subsequently, w will be transformed and passed to the synthesis network to better assist in learning the features of the fingerprint. Synthesis Network: It consists of multiple sub-networks trained progressively. The input of the first sub-network is a constant, and the input of the remaining sub-networks is the output of the previous sub-network. The first sub-network includes 1 convolutional layer, and the remaining sub-networks include 2 convolutional layers and an upsampling layer. Module A and Module B are added to each layer of the sub-network. Module A is the transformed random noise, which is used to improve the robustness of the model. Module B is the affine transformation obtained by transforming the intermediate vector w after feature disentanglement, which is used to assist the synthesis network in better learning the features of the fingerprint. Each dimensional level of the synthesis network will be affected by Module B twice. One time is applied after upsampling (since the first sub-network does not have an upsampling layer, so the influence of Module B is directly applied after the input constant), and the other time is applied after convolution. The influence method uses BN (batch normalization).

[0062] The task of discriminator D is to distinguish the real data Ξ and the generated data G(z, Ψ). By adjusting the parameters, it makes the value of D(Ξ, Ψ) as close to 1 as possible for the real data, and the value of D(G(z, Ψ), Ψ) as close to 0 as possible for the generated data. Here, D(Ξ, Ψ) is the output of the discriminator for the real data Ξ and the condition Ψ, and D(G(z, Ψ), Ψ) is the output of the discriminator for the generated data G(z, Ψ) and the condition Ψ. The corresponding loss function is where the first term represents the expected loss that the discriminator judges the real data as true. The discriminator hopes that the output of D(Ξ, Ψ) for the real data is as close to 1 as possible. At this time, log D(Ξ, Ψ) is close to 0 and the loss is the smallest. The second term represents the expected loss that the discriminator judges the generated data as false. The discriminator hopes that the output of D(G(z, Ψ), Ψ) for the generated data is as close to 0 as possible. At this time, log(1 - D(G(z, Ψ), Ψ)) is close to 0 and the loss is the smallest. The discriminator calculates the gradient according to the loss function and updates the parameters to improve its ability to distinguish real data and generated data.

[0063] We compared the performance of the positioning method disclosed in the embodiments of the present invention with other positioning methods through simulation. We used a geometric-based 2D propagation model to simulate the wireless transmission environment. Assume that three base stations are located at (0m, 200m), At each location, a uniform linear array is equipped. Multi-base station cooperation can achieve positioning in a large area. Assume that there are several scatterers randomly distributed in the area to be located. For simplicity without losing generality, only the single-reflection propagation of wireless signals by scatterers is considered. To realistically restore the complete scatterer environment of the area to be located, all scatterers in the area to be located are included in the calculation in the embodiments of the present invention. To demonstrate the performance of this positioning method, the estimated position coordinates of the mobile terminal to be located are compared with its true position coordinates, and the cumulative distribution function (CDF) of the positioning error is calculated. The simulation parameters are shown in Table 1.

[0064] Table 1 Simulation Parameters

[0065]

[0066] Figure 4 The positioning accuracies of the positioning methods using ADCPM fingerprints and ADSM fingerprints at different reference point intervals are compared. The models both use DCNNC, the number of antennas on the base station side is 128, and the scatterer density is 0.020 / m 2 . It can be seen that regardless of whether the reference point interval is 2 m or 4 m, the positioning accuracy of using ADSM fingerprints is very close to that of using ADCPM fingerprints. And when the reference point intervals are the same, the storage overhead of ADCPM fingerprints is much larger than that of ADSM fingerprints.

[0067] Figure 5 The comparison is made when the number of antennas on the base station side is 128, the reference point interval is 2 m, and the scatterer density is 0.020 / m 2The positioning accuracies of the multi-base station joint positioning method based on FSGAN, the single-base station positioning method based on FSGAN, the single-base station positioning method based on deep convolutional neural network classification (DCNNC), the single-base station positioning method based on deep convolutional neural network regression (DCNNR), and the single-base station positioning method based on two-stage fingerprint clustering (TSFC) adopted in the embodiments of the present invention all use ADSM fingerprints. The multi-base station joint positioning method based on FSGAN provides a 1m positioning accuracy with a 95% confidence probability, offering the highest performance; the single-base station positioning method based on FSGAN provides a 1m positioning accuracy with an 80% confidence probability; the positioning method based on DCNNC can provide a 1m positioning accuracy with a 72% confidence probability; the positioning method based on DCNNR can provide a 1m positioning accuracy with a 53% confidence probability, while the positioning method based on TSFC only has a 20% confidence probability for a 1m positioning accuracy. It can be seen that the multi-base station joint positioning method based on FSGAN adopted in the embodiments of the present invention has an absolute advantage in terms of positioning accuracy. This is because the multi-base station joint positioning comprehensively utilizes the features extracted by each single base station. Through the constructed joint fingerprint, the features extracted by each single base station restrict each other, and higher-precision positioning of users is achieved during the mutual restriction process. In addition, the performance of the FSGAN model is superior to other methods because the adversarial training mechanism of FSGAN can provide more accurate regression results.

[0068] The embodiments of the present invention also disclose a computer system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the steps of the network large-scale MIMO multi-base station joint positioning method disclosed in the above embodiments are implemented.

[0069] The embodiments of the present invention also disclose a computer program product, including a computer program. When the computer program is executed by the processor, the steps of the network large-scale MIMO multi-base station joint positioning method disclosed in the above embodiments are implemented.

[0070] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A network large-scale MIMO multi-base station joint positioning method, characterized in that: The following steps are involved: Each base station extracts a low-dimensional angle delay domain sparse matrix (ADSM) as a location fingerprint based on the statistical channel information, and uses the ADSM fingerprint and the corresponding location coordinates as a data set to train the first network model and preliminarily extract features; The features extracted by all single base stations are fused into a joint fingerprint, and the joint fingerprint and the corresponding location coordinates are used as a data set to train the second network model. For the user to be located, a joint fingerprint is constructed based on the features extracted by all single base stations through the first network model, and the user location is estimated using the trained second network model.

2. The network massive MIMO multi-base station joint positioning method according to claim 1, characterized in that: The angle delay domain sparse matrix (ADSM) It is the dimension reduction transformation of the angle delay domain channel energy matrix Ω, where N s Represents the number of peak elements in the original Ω matrix; the three columns of Ψ represent the row index, column index, and the value of the element at the corresponding row and column index position of the peak element in the Ω matrix.

3. The network massive MIMO multi-base station joint positioning method according to claim 1, characterized in that: The first network model is a feature separation generative adversarial network (FSGAN), which includes a generator G and a discriminator D. The generator includes a mapping network and a synthesis network. FSGAN combines GAN with supervised learning by adding labels and uses conditional information to guide the generation process. In the location prediction task, the fingerprint condition Ψ related to the location Ξ is used as input, and the distribution p of the real data is learned. data (Ξ|Ψ) and generate similar but not identical data p g (Ξ|Ψ).

4. The network massive MIMO multi-base station joint positioning method according to claim 3, characterized in that: The FSGAN adopts progressive training to capture the potential relationship between location data and related fingerprints; the synthetic network contains multiple sub-networks, each with different sample dimensions, which are gradually improved to refine the learned features.

5. The network massive MIMO multi-base station joint positioning method according to claim 4, characterized in that: During the training process, the generator and discriminator are first trained using low-dimensional noise samples, and then each time they are upgraded to a higher-dimensional sub-network; during the dimensionality increase process, a mixing coefficient α is introduced for a smooth transition.

6. The network massive MIMO multi-base station joint positioning method according to claim 3, characterized in that: The mapping network consists of multiple fully connected layers, which are used to decouple the input fingerprint into an intermediate hidden variable w; the synthesis network consists of multi-layer sub-networks trained progressively, wherein the input of the first layer of sub-network is a constant, and the input of the remaining sub-networks is the output of the previous layer of sub-networks; the first layer of sub-network includes a convolutional layer, and the remaining sub-networks include two convolutional layers and an upsampling layer; each layer of sub-network is added with converted random noise and an affine transformation of w; each dimensional level of the synthesis network will be affected by the affine transformation of w, and the influence method adopts batch normalization.

7. The network massive MIMO multi-base station joint positioning method according to claim 3, characterized in that: The second network model is defined as a joint FSGAN model, and its structure is similar to the FSGAN of the first network model.

8. The network massive MIMO multi-base station joint positioning method according to claim 1, characterized in that: The joint fingerprint information and position coordinates of the reference points are used as training data to train the second network model, and the trained network structure is stored in the database. When multiple base stations are jointly positioned, each base station extracts the ADSM position fingerprint information from the channel state information obtained from the uplink channel estimation, and uses the first network model trained by itself to extract the position features of the terminal to be positioned; The location features extracted by each base station for the terminal to be located are integrated into a joint fingerprint of the terminal to be located, which is input into the trained second network model to output the final predicted location of the terminal to be located.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the network large-scale MIMO multi-base station joint positioning method according to any one of claims 1-8 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the network large-scale MIMO multi-base station joint positioning method according to any one of claims 1-8 are implemented.

Images