5GNR indoor positioning method, program and system based on dynamic multi-graph aggregation network, and storage medium

Abstract

The invention discloses a 5G NR indoor positioning method, program and system based on a dynamic multi-graph aggregation network, and a storage medium, and belongs to the field of deep learning. Three-dimensional space coordinates are obtained by preprocessing CSI data, constructing a CSI graph structure, establishing a spatial relationship in CSI, performing feature aggregation by using a graph convolutional network GCN, performing fusion of local and global features, extracting features of a CSI graph, dynamically aggregating CSI sub-graphs, and performing position mapping. According to the method, multi-channel feature information is utilized, topological connection in a non-Euclidean space is introduced, data is expressed as graphic information, therefore, the spatial relation between different channels is analyzed more comprehensively, meanwhile, a dynamic sub-graph aggregation mode is used, and finally higher-precision positioning application is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of computers, and in particular to the field of deep learning in computers, and specifically to a 5GNR indoor positioning method, program, system and storage medium based on a dynamic multi-graph aggregation network. Background Art

[0002] Indoor positioning technology based on channel state information (CSI) has received widespread attention in recent years, especially against the backdrop of the rapid development of wireless communication technology. Traditional positioning methods, such as the Global Navigation Satellite System (GNSS), face signal blocking and attenuation issues in indoor environments, resulting in a significant decrease in positioning accuracy. Therefore, new solutions are urgently needed for indoor positioning technology. Wireless communication technologies, particularly WiFi, Bluetooth, ultra-wideband (UWB), and 5G, have provided new opportunities for indoor positioning. WiFi and Bluetooth have been widely used due to their low cost and high penetration rate. However, they primarily rely on received signal strength indication (RSSI) for positioning, which is susceptible to environmental interference, resulting in unstable positioning accuracy. UWB has strong anti-interference capabilities, but it suffers from signal attenuation when encountering obstacles. With the advent of 5G technology, its advantages of high bandwidth, low latency, and large capacity have enabled significant progress in indoor positioning technology. However, the complexity of 5G signals and issues such as signal interference make it more difficult to directly extract effective feature information from 5G signals. Recent research has proposed deep learning methods for scaling graph data. Among them, graph convolutional networks (GCNs) have demonstrated their effectiveness in graph data applications. By increasing the number of layers, GCN can aggregate node information and focus not only on local details but also on global details. However, multi-layer graph convolution may lead to over-smoothing and too many parameters. Summary of the Invention

[0003] To address these issues, this paper proposes a 5GNR indoor positioning method based on a dynamic multi-graph aggregation network. By leveraging multi-channel feature information and introducing topological connections in non-Euclidean space, the data is represented as graph information, enabling a more comprehensive analysis of the spatial relationships between different channels. Furthermore, a dynamic subgraph aggregation method is employed to achieve higher-precision positioning applications.

[0004] The present invention provides a 5GNR indoor positioning method based on a dynamic multi-graph aggregation network:

[0005] A 5GNR indoor positioning method based on a dynamic multi-graph aggregation network sets up five 5G base stations and five signal receiving devices indoors, and obtains CSI data through cellular positioning technology;

[0006] By preprocessing the CSI data, a CSI graph structure is constructed;

[0007] By inputting CSI graph information into the GCN model, relevant information of different patches is aggregated to obtain CSI subgraphs;

[0008] By taking each CSI subgraph as input, we dynamically aggregate the CSI subgraphs and transform them into observable real-valued form;

[0009] The three-dimensional spatial coordinates (X, Y, Z) are obtained by performing average pooling and nonlinear mapping on the observable real-valued form set.

[0010] Furthermore, the CSI data is preprocessed to construct a CSI graph structure as follows:

[0011] Step 1.1: Use filters to capture channel information differences and extract channel features through grouped convolutional layers to downsample the CSI data X to obtain node units X out ;

[0012]

[0013] Where: C represents the convolution layer, BN is the batch normalization of the convolution data, σ is the ReLU activation function, g in and g out are the input and output dimensions, k w×w Indicates the size of the convolution kernel;

[0014] Step 1.2: Set each node element X out Considered as a feature vector x i , for each eigenvector x i Add position code i , obtain the node unit set;

[0015] x i +e i →x i

[0016] X out =[x1,x2,x3,…,x N ]

[0017] in: D is the feature dimension, N is the number of small blocks, and for each small block, we add a position encoding vector e i ,

[0018] Step 1.3: Treat the node unit set as an unordered set of nodes. For each node, use the Euclidean distance to find its K nearest neighbors and add edges. Add edge information e for each edge.ij , build graph structure

[0019]

[0020] Where: (i) Representation diagram The edge set of Representation diagram The node set of , K is a set constant.

[0021] Furthermore, the GCN model includes a graph convolution layer, a ReLU activation layer and a Dropout layer.

[0022] Furthermore, the CSI graph information is input into the GCN model, and the relevant information of different patches is aggregated to obtain the CSI subgraph as follows:

[0023] Step 2.1: Input the CSI graph structure into the graph convolution layer, perform node feature aggregation, and update it using the feature information of adjacent nodes;

[0024] The polymerization process is as follows:

[0025]

[0026] in: represents the aggregation operation of the k-th layer graph convolution node, Represents node v i The feature information at the kth layer is updated iteratively with the number of layers k; Agg(·) and Update(·) represent the aggregation and update operations of the graph, respectively; Represents node v i and v j The edge information between them, N(V) is the set of all nodes in this graph;

[0027] Step 2.2: After performing the graph convolution operation, use the ReLU activation function to introduce nonlinear factors and introduce the feedforward neural network FFN;

[0028] The specific operations of FFN are as follows:

[0029]

[0030] Among them: FC1 and FC2 are fully connected layers with batch normalization operations; ReLU is the activation function; Dropout is a regularization operation to reduce overfitting;

[0031] Step 2.3: Extract local feature aggregation and global feature aggregation through two layers of graph convolutional network, fuse local and global features, and obtain subgraph Y.

[0032] Y=[y1,y2,y3,…,y j ];

[0033] Furthermore, the method of taking each CSI sub-graph as input, dynamically aggregating the CSI sub-graphs, and converting them into an observable real-valued form is specifically as follows:

[0034] Step 3.1: Each subgraph y j As input, wave information representing amplitude and phase information;

[0035]

[0036] in: is the subgraph y j The wave representation of i is the imaginary unit; ⊙ represents the Hadamard product; ||z j || is a real-valued feature of each subgraph; is a periodic function, θ j is the phase, which indicates the current position of the subgraph in the waveform cycle;

[0037] For amplitude information, the real-valued features of the subgraph are extracted through the channel-FC operation:

[0038]

[0039] in: is a learnable parameter, y j represents the jth subgraph;

[0040] For phase information, dynamic phase information is generated by learning parameters:

[0041]

[0042] in: is a learnable parameter, y j represents the jth subgraph;

[0043] Step 3.2: Expand the complex representation of the subgraph into separate representations of the real and imaginary parts, and aggregate the subgraph information through a multi-layer perceptron (MLP).

[0044]

[0045] Where: ||z r || is the mixed real-valued feature, i and j represent different subgraphs, ||z i ||、||z j || are the real-valued features of the i-th and j-th subgraphs, θ i ,θ j are the phase information of the i-th and j-th sub-images respectively;

[0046] Step 3.3: Convert the complex representation of the aggregate into an observable real-valued form:

[0047]

[0048] Where: t and ω i are learnable weights, ||z r || is the mixed real-valued feature, θ r is the phase information after mixing.

[0049] Furthermore, the observable real-valued form set is average pooled and nonlinearly mapped to obtain the three-dimensional space coordinates (X, Y, Z) as follows:

[0050] Step 4.1: Perform a global average pooling operation on the extracted features, compress the spatial dimension to 1×1 to capture the global semantic representation, and obtain the aggregated features

[0051]

[0052] Among them: o is the observable real-valued form set obtained;

[0053] Step 4.2: Through the convolution layer, batch normalization, activation function and fully connected layer, the aggregated features are transformed into Map to three-dimensional space coordinates (X, Y, Z);

[0054]

[0055] The present invention also provides a computer device / equipment / system, comprising a memory, a processor, and a computer program stored on the memory, wherein when the processor executes the computer program, the steps of the 5G NR indoor positioning method based on a dynamic multi-graph aggregation network as described in any one of the above items are implemented.

[0056] The present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of any of the above-mentioned 5G NR indoor positioning methods based on a dynamic multi-graph aggregation network.

[0057] The present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of any of the above-mentioned 5G NR indoor positioning methods based on a dynamic multi-graph aggregation network.

[0058] The beneficial effects of the present invention are:

[0059] By introducing a feed-forward network (FFN) into graph convolution and extracting semantics from graph structures at different levels, we preserve a wealth of feature information. Furthermore, we propose a method for dynamically learning subgraph aggregation weight coefficients. This method not only considers the diversity of subgraphs but also captures a wider range of feature information through subgraph aggregation. We demonstrate that leveraging semantics from different levels of graph structures and dynamic multi-graph aggregation are more effective than traditional graph convolution methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Establish a spatial relationship execution flow chart in CSI;

[0061] Figure 2 Flowchart of feature extraction execution of CSI graph;

[0062] Figure 3 Dynamic aggregation CSI subgraph execution flow chart;

[0063] Figure 4 Ablation experiment diagram;

[0064] Figure 5 Flowchart of overall steps. DETAILED DESCRIPTION

[0065] The present invention includes three main modules: establishing spatial relationships in CSI, feature extraction of CSI graphs, and GCN processing and subgraph aggregation.

[0066] First, let's introduce the CSI data used. This invention is based on cellular positioning technology and uses data from the 3rd Generation Partnership Project Release 16 (3GPP R16) standard. Orthogonal frequency division multiplexing (OFDM) demodulation and time offset correction are performed when receiving the signal, allowing each user equipment (UE) to receive the resource grid. The subcarriers for each antenna are represented as a column vector of dimension k, as shown in Equation 1:

[0067]

[0068] Where k represents the number of subcarriers, H raw represents the channel impulse response, refGrid represents the known pilot, and rxGrid represents the received resource grid.

[0069] To ensure signal integrity and improve the accuracy of spectrum analysis, the authors performed a frequency domain analysis of the original channel frequency response H raw The interpolation process is performed on both sides of . Subsequently, the information is converted into a channel impulse response (CIR) using an inverse Fourier transform. The denoised channel impulse response is expressed as:

[0070] H denoised =H raw W

[0071] Where W is the time-domain raised cosine window used for CIR denoising.

[0072] Finally, the denoised channel impulse response is converted back to the time domain through inverse Fourier transform, so as to understand the channel characteristics more intuitively:

[0073]

[0074] Finally, noise with different signal-to-noise ratio (SNR) levels was added to the processed CSI data to generate three datasets with different SNRs: SNR10, SNR20, and SNR50. Specifically, each dataset contains 4816 positioning samples, and the information contained in each sample can be expressed as:

[0075]

[0076] Where m represents the number of MIMO channels and n represents the number of subcarriers per antenna. In this study, the UE receives signals from five different base stations.

[0077] To demonstrate the effectiveness of the present invention, the following experimental demonstration was conducted. The experimental data consists of three real-world 5G scenario datasets collected from the indoor space of the new laboratory building of the Chinese Academy of Sciences in Beijing. The entire indoor space is 20 meters wide, 60 meters long, and 4 meters high, making it a large room suitable for applications such as factories and museums. To obtain CSI, five 5G base stations were deployed using integrated sensing and communication at 3.5 GHz, with a bandwidth of 100 MHz and a power of 40 watts. These base stations were mounted on plastic supports at a height of 2.4 meters above the ground, and a random floating height of 0.1 meters was introduced during the simulation to prevent coplanarity. The user device acted as a receiver and was placed on a marked small car at a height of 1.2 meters above the ground, simulating a person holding a mobile phone at a height of 1.8 meters. The obtained dataset includes 4816 positioning samples, with three datasets corresponding to different representations of CSI at SNR (signal-to-noise ratio) of 10, SNR20, and SNR50. To split the dataset into training, validation, and test sets, we used an approximate 6:2:2 ratio based on the signal-to-noise ratio (SNR), resulting in 2888, 964, and 964 samples, respectively. The size of a single CSI matrix is ​​5 × 16 × 193, indicating the presence of five base stations, each with 16 antennas and 193 subcarriers.

[0078] The initial learning rate is 5×10 -3, using Adam optimizer, training epoch 300, batch size batch_size 16.

[0079] The two evaluation criteria used in the experiment are Root Mean Square Error (RMSE) and Mean Absolute Error (MAE):

[0080]

[0081] Among them S i Indicates its actual location, represents the predicted location of the i-th user equipment, and N is the total number of user equipment.

[0082] The comparison methods of the present invention include: CLnet ("Clnet: Complex input lightweight neural network des igned for massive MIMO CSI feedback" https: / / ieeexplore.ieee.org / abstract / document / 9497358), Crnet ("Multi-resolution CSI feedback with deeplearning in massive mimo" system》https: / / ieeexplore.ieee.org / document / 9149229), KNN, CSInet(《Deep learning for massive MIMO CSI feedback》https: / / ieeexplore.ieee.org / abstract / document / 8322184), Hiloc(《Hi-Loc:Hybrid IndoorLocalization via Enhanced 5G NR CSI》https: / / ieeexplore.ieee.org / abstra ct / document / 9855525), SVM, MPRI (《Toward 5G NR High-Precision Indoor Positioningvia Channel Frequency Response: A New Paradigm and Dataset Generation Method》https: / / ieeexplore.ieee.org / abstract / document / 9729785), SSQ-Net(《A dl-basedhigh-precision posi tioning method in challenging urban scenarios for b5gccuavs》https: / / ieeexplore.ieee.org / doc ument / 10130727), E-NE(《Vision gnn: Animage is worth graph of nodes》https: / / arxiv.or g / abs / 2206.00272).

[0083] The experiment compared the error between the present invention and other methods, as shown in Table 1.

[0084] Table 1 Comparison of error sizes of different methods

[0085]

[0086]

[0087] By analyzing Table 1, we can draw the following conclusions:

[0088] 1) Compared with SVM, KNN, CLnet, CRnet, CSInet, SSQ-Net, MPRI, Hiloc, and E-NE methods, the present invention significantly reduces the error size and effectively improves the prediction accuracy.

[0089] The ablation experiment results are as follows Figure 4 As shown, analysis Figure 4 The results demonstrate the key role of the multi-scale sub-image module and the dynamic sub-image aggregation module in improving localization performance. These two modules complement each other, contributing to the DGA model's efficient feature extraction and fusion capabilities. In the figure, DGA-N1 is a model without the multi-scale sub-image module, and DGA-N2 is a model without the dynamic sub-image aggregation module.

[0090] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution in the embodiment of the present invention will be supplemented below with reference to the drawings in the embodiment of the present invention, so that the solution can be described more clearly and completely.

[0091] The four steps are CSI and sequence type data conversion operation, subcarrier modeling operation in a single channel (single base station), subcarrier modeling operation in multiple channels (multiple base stations), and position mapping operation.

[0092] The details are as follows:

[0093] Step S1: Establish spatial relationship in CSI:

[0094] Step S1.1 CSI data preprocessing:

[0095] A 3×3 filter is used to capture the channel information difference, and the channel features are extracted through the grouped convolution layer to downsample the CSI data X to obtain the node unit X. out ;

[0096] Each node element X out Considered as a feature vector x i , for each eigenvector x i Add position code i , get the node unit set.

[0097] Step S1.2: Construct the CSI graph structure:

[0098] Treat the extracted features as an unordered set of nodes Use Euclidean distance to find the K nearest neighbors of each node and add edge information e for each edge ij . Constructing the graph structure

[0099]

[0100] Step S2: Feature extraction of CSI image:

[0101] Step S2.1 uses graph convolutional network GCN for feature aggregation:

[0102] The CSI graph data is input and graph convolution is performed to aggregate the relevant information of different small blocks, further extract node features and update their representations. Each GCN layer consists of a graph convolution layer, a ReLU activation layer, and a Dropout layer, using residual connections to retain key information.

[0103] Step S2.2: Fusion of local and global features:

[0104] Local features and global features are extracted respectively through a two-layer graph convolutional network, and then fused to obtain the CSI subgraph.

[0105] Step S3 dynamically aggregates CSI subgraphs:

[0106] By dynamically aggregating CSI subgraphs to represent amplitude and phase information, the feature representation is performed in the complex domain. Subgraph features are aggregated using an MLP. Subgraph information is aggregated through dynamic phase adjustment to obtain a complex representation, which is then converted to an observable real-valued form.

[0107] Step S4: Position Mapping:

[0108] A global average pooling operation is performed on the extracted features to compress the spatial dimension to 1×1 to capture global semantic representation. The aggregated features are mapped to three-dimensional spatial coordinates (X, Y, Z) through convolutional layers, batch normalization, activation functions, and fully connected layers.

[0109] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A 5GNR indoor positioning method based on a dynamic multi-graph aggregation network, characterized by: Five 5G base stations and five signal receiving devices were set up indoors to obtain CSI data using cellular positioning technology; By preprocessing the CSI data, a CSI graph structure is constructed; By inputting CSI graph information into the GCN model, relevant information of different patches is aggregated to obtain CSI subgraphs; By taking each CSI subgraph as input, we dynamically aggregate the CSI subgraphs and transform them into observable real-valued form; The three-dimensional spatial coordinates (X, Y, Z) are obtained by performing average pooling and nonlinear mapping on the observable real-valued form set.

2. A 5G NR indoor positioning method based on a dynamic multi-graph aggregation network according to claim 1, characterized in that: The CSI data is preprocessed to construct the CSI graph structure as follows: Step 1.1: Use filters to capture channel information differences and extract channel features through grouped convolutional layers to downsample the CSI data X to obtain node units X out ; Where: C represents the convolution layer, BN is the batch normalization of the convolution data, σ is the ReLU activation function, g in and g out are the input and output dimensions, k w×ω Indicates the size of the convolution kernel; Step 1.2: Set each node element X out Considered as a feature vector x i , for each eigenvector x i Add position code i , obtain the node unit set; x i +e i →x i X out =[x1,x2,x3,…,x N ] in: D is the feature dimension, N is the number of small blocks, and for each small block, we add a position encoding vector e i , Step 1.3: Treat the node unit set as an unordered set of nodes. For each node, use the Euclidean distance to find its K nearest neighbors and add edges. Add edge information e for each edge. ij , build graph structure Where: ε (i) Representation diagram The edge set of Representation diagram The node set of , K is a set constant.

3. The 5G NR indoor positioning method based on a dynamic multi-graph aggregation network according to claim 1, characterized in that: The GCN model includes a graph convolution layer, a ReLU activation layer, and a Dropout layer.

4. The 5G NR indoor positioning method based on a dynamic multi-graph aggregation network according to claim 3, characterized in that: The CSI graph information is input into the GCN model, and the relevant information of different patches is aggregated to obtain the CSI subgraph as follows: Step 2.1: Input the CSI graph structure into the graph convolution layer, perform node feature aggregation, and update it with the feature information of adjacent nodes; The polymerization process is as follows: in: represents the aggregation operation of the k-th layer graph convolution node, Represents node v i The feature information at the kth layer is updated iteratively with the number of layers k; Agg(·) and Update(·) represent the aggregation and update operations of the graph, respectively; Represents node v i and v j The edge information between them, N(V) is the set of all nodes in this graph; Step 2.2: After performing the graph convolution operation, use the ReLU activation function to introduce nonlinear factors and introduce the feedforward neural network FFN; The specific operations of FFN are as follows: Among them: FC1 and FC2 are fully connected layers with batch normalization operations; ReLU is the activation function; Dropout is a regularization operation to reduce overfitting; Step 2.3: Extract local feature aggregation and global feature aggregation through two layers of graph convolutional network, fuse local and global features, and obtain subgraph Y. <h2 style=";text-align:left;direction:ltr">Y = [y1,y2,y3,…,y<h2 style=";text-align:left;direction:ltr"> j <h2 style=";text-align:left;direction:ltr"> ]。 5. The 5GNR indoor positioning method based on a dynamic multi-graph aggregation network according to claim 4, characterized in that: The process of taking each CSI subgraph as input, dynamically aggregating the CSI subgraphs, and converting them into an observable real-valued form is as follows: Step 3.1: Each subgraph y j As input, wave information representing amplitude and phase information; in: is the subgraph y j The wave representation of i is the imaginary unit; ⊙ represents the Hadamard product; ||z j || is a real-valued feature of each subgraph; is a periodic function, θ j is the phase, which indicates the current position of the subgraph in the waveform cycle; For amplitude information, the real-valued features of the subgraph are extracted through the channel-FC operation: in: is a learnable parameter, y j represents the jth subgraph; For phase information, dynamic phase information is generated by learning parameters: in: is a learnable parameter, y j represents the jth subgraph; Step 3.2: Expand the complex representation of the subgraph into separate representations of the real and imaginary parts, and aggregate the subgraph information through a multi-layer perceptron (MLP). Where: ||z r || is the mixed real-valued feature, i and j represent different subgraphs, ||z i ||、||z j || are the real-valued features of the i-th and j-th subgraphs, θ i ,θ j are the phase information of the i-th and j-th sub-images respectively; Step 3.3: Convert the complex representation of the aggregate into an observable real-valued form: Where: t and ω i are learnable weights, ||z r || is the mixed real-valued feature, θ r is the phase information after mixing.

6. The 5G NR indoor positioning method based on a dynamic multi-graph aggregation network according to claim 5, characterized in that: The observable real-valued form set is average pooled and nonlinearly mapped to obtain the three-dimensional space coordinates (X, Y, Z) as follows: Step 4.1: Perform a global average pooling operation on the extracted features, compress the spatial dimension to 1×1 to capture the global semantic representation, and obtain the aggregated features Among them: o is the observable real-valued form set obtained; Step 4.2: Through the convolution layer, batch normalization, activation function and fully connected layer, the aggregated features are transformed into Map to three-dimensional space coordinates (X, Y, Z); 7. A computer device / apparatus / system comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer program product comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images