A CSI indoor positioning method and device based on a graph neural network

Abstract

The application discloses a CSI indoor positioning method and device based on a graph neural network, and belongs to the technical field of indoor positioning. The method comprises the following steps: arranging a plurality of access points in an indoor positioning scene, and arranging a plurality of reference points in a positioning area; acquiring CSI fingerprint features between a user terminal and the plurality of access points at different reference points, and extracting amplitude features and phase features as node features from the CSI fingerprint features; obtaining edge features and an adjacency matrix according to topological information between the access points, completing construction of a graph structure, and then obtaining a graph dataset; using the graph dataset to train a CSI-GNN offline until a loss function is minimized or a set number of iterations is reached; and using a trained CSI-GNN algorithm model for online positioning, and outputting a real-time position of the user terminal. The CSI-GNN algorithm disclosed by the application can dynamically evaluate and optimize the relative importance of amplitude information and phase information, thereby improving the accuracy and stability of indoor positioning.

Description

Technical Field

[0001] This invention belongs to the field of indoor positioning technology, specifically relating to a CSI indoor positioning method and device based on graph neural networks. Background Technology

[0002] In the field of WiFi indoor positioning, with the widespread deployment of wireless networks and the popularity of smart devices, various WiFi-based positioning methods have become a hot topic in research and application. Received Signal Strength Indicator (RSSI) technology is one of the most commonly used methods for WiFi indoor positioning. It estimates the location of the user device based on the signal strength received from different access points. However, in complex indoor environments, its positioning accuracy is significantly reduced by physical obstacles and multipath effects.

[0003] Channel State Information (CSI) indoor positioning technology is a more advanced wireless positioning technology that better copes with multipath effects and environmental changes, thus achieving higher positioning accuracy. CSI contains detailed propagation information of the wireless signal on each subcarrier, including amplitude attenuation and phase shift. This information reflects multipath effects such as reflection, diffraction, and scattering of the signal when it encounters obstacles. Because CSI can capture subtle signal changes, it can be used to sense environmental changes, such as changes in the user's location in an indoor environment.

[0004] A common CSI-based indoor positioning method is to create a CSI fingerprint map, which involves collecting CSI data at different locations within the indoor environment. During positioning, the real-time collected CSI data is matched with data in the fingerprint database to determine the current location.

[0005] Patent document CN110109055A discloses an indoor positioning method based on RSSI ranging, comprising: arranging substation nodes in a square topology, using ZigBee technology for networking and communication, tag nodes broadcasting data and sending tag data packets, which are then parsed by the substation nodes, reassembled into substation data packets, and sent to the base station node. The base station node extracts the RSSI value from the data packets, calculates the distance between the tag nodes and each substation node, and optimizes the distance values ​​multiple times using Kalman filtering and iterative methods to eliminate the influence of external interference on the accuracy of the distance values, thereby ensuring positioning accuracy. However, the RSSI used in this invention is greatly affected by signal attenuation and multipath effects, which may lead to unstable signal strength, thus limiting positioning accuracy. In addition, there are many factors in the indoor environment that may interfere with the signal, such as walls and furniture, which can also affect signal propagation and reception, thereby affecting the accuracy of RSSI positioning.

[0006] Patent document CN111212379A discloses a novel CSI indoor positioning method based on convolutional neural networks (CNNs). The method includes: acquiring CSI data and extracting three types of data features: average amplitude, phase difference, and CIR amplitude distribution central moment; constructing an image from the data features and using a CNN to map location information to the feature image; training the CNN using a dataset composed of images and category labels; and performing online positioning using the trained CNN, achieving location estimation through a multi-image positioning algorithm based on spectral clustering. However, CNN-based CSI indoor positioning technology does not utilize the topological information between access points; it simply stitches together CSI features from different access points into an image and then inputs it into the model for training and prediction. This approach leads to the loss of access point topological information, thus limiting positioning accuracy. Summary of the Invention

[0007] The purpose of this invention is to provide a CSI indoor positioning method and device based on graph neural networks. Indoor positioning based on CSI can fully utilize the amplitude and phase information of CSI while retaining the access point topology information. The proposed CSI-GNN algorithm can dynamically evaluate and optimize the relative importance of amplitude and phase information, enabling the method to effectively adapt to different indoor environmental conditions, thereby improving the accuracy and stability of the indoor positioning system.

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0009] In a first aspect, the present invention provides a CSI indoor positioning method based on graph neural networks, comprising the following steps:

[0010] Step 1: Construct an indoor positioning scenario, including: setting up several access points for transmitting signals in the indoor positioning scenario, setting up a positioning area in the indoor positioning scenario, and setting up several reference points in the positioning area to represent the location of the user terminal.

[0011] Step 2: The user terminal receives transmitted signals from several access points at different reference points and extracts CSI fingerprint features respectively. The CSI fingerprint features are dimensionality reduced to obtain node features, which include amplitude features and phase features. Edge features and adjacency matrix are obtained based on the topological information between access points. The graph structure corresponding to the CSI fingerprint features is constructed based on the node features, edge features and adjacency matrix and the graph category label is labeled. The set of graph structure and its corresponding graph category label is the graph dataset.

[0012] Step 3: Train the CSI-GNN algorithm model offline using the graph dataset, which includes GNN1, GNN2 and MLP. This includes: inputting amplitude features into GNN1, phase features into GNN2, edge features and adjacency matrix into GNN1 and GNN2, concatenating the output features of GNN1 and GNN2 and inputting them into MLP, dynamically adjusting the weight parameters of MLP, outputting the prediction results, and minimizing the loss function between the prediction results and the graph category labels, or reaching the maximum number of iterations, to obtain the trained CSI-GNN algorithm model.

[0013] Step 4: Use the trained CSI-GNN algorithm model to perform online localization and output the real-time location of the user terminal.

[0014] The technical concept of this invention is as follows: construct an indoor positioning scenario that includes access points and user terminal location reference points; perform indoor positioning based on CSI; when the user terminal is located at any reference point, the user terminal receives the transmitted signals from all access points in the indoor positioning scenario, extracts the CSI fingerprint features from them, performs dimensionality reduction processing, obtains the amplitude information and phase information of the transmitted signals respectively, and constructs feature vectors containing these two types of information, namely amplitude features and phase features, as node features of the corresponding access points.

[0015] Meanwhile, this invention further considers the importance of topological information between various access points in indoor scenes for positioning accuracy. Therefore, this invention obtains edge weights and adjacency matrices based on the topological information between access points. The edge weights serve as edge features, and the adjacency matrix is ​​used to represent the adjacency relationships of nodes in the graph structure. A graph structure is constructed based on node features, edge features, and the adjacency matrix, and the corresponding graph category labels are obtained. These graph category labels are used to characterize the reference point where the user terminal is located. The set of graph structures and their corresponding graph category labels constitutes a graph dataset.

[0016] Next, a CSI-GNN algorithm model is trained using a graph dataset. The CSI-GNN algorithm model is trained by minimizing the cross-entropy loss function. The trained CSI-GNN algorithm model is then used for actual inference.

[0017] Furthermore, in step 2, the CSI fingerprint feature is expressed by the formula:

[0018]

[0019] Among them, CSI AP This represents the data obtained from N at a certain reference point. AP A set of CSI fingerprint features for each access point, CSI jIt is an N×M×K matrix representing the CSI fingerprint features between the user terminal at the current reference point and the j-th access point, where j = 1, 2, ..., N. AP N represents the number of receiving antennas for the user terminal, M represents the number of transmitting antennas for each access point, and K represents the number of subcarriers.

[0020] Furthermore, in step 2, the dimensionality reduction processing of the CSI fingerprint features to obtain node features, which include amplitude features and phase features, specifically:

[0021] For CSI j Dimensionality reduction is performed to obtain a k-dimensional vector, which can be expressed by the formula:

[0022]

[0023] Among them, L j Let l represent the k-dimensional vector corresponding to the j-th node. jt L represents j The magnitude of the t-th element, θ jt L represents j The phase of the t-th element;

[0024] The amplitude and phase features are extracted from the k-dimensional vector, expressed by the following formula:

[0025]

[0026]

[0027] Where LA represents the amplitude characteristic, LP represents the phase characteristic, and la j =[l j1 ,l j2 ,...,l jk ], corresponding to CSI j Amplitude information, lp j =[θ j1 ,θ j2 ,...,θ jk ], corresponding to CSI j Phase information.

[0028] Furthermore, in step 2, obtaining edge features based on the topological information between access points specifically involves:

[0029] The topology information is based on the physical distance between access points;

[0030] The edge weights are obtained based on the physical distance between access points, and can be expressed by the following formula:

[0031]

[0032] Among them, w ij Denotes edge weight, d ij This represents the physical distance between the i-th access point and the j-th access point;

[0033] Edge weights are used as edge features.

[0034] Furthermore, in step 2, based on the obtained edge weights, the topology information uses the adjacency relationships between access points to construct an adjacency matrix, which is expressed by the formula:

[0035]

[0036] Where A represents the adjacency matrix, a ij This represents the edge feature from the i-th node to the j-th node. When i ≠ j, a ij =w ij When i = j, a ij =a ii =0.

[0037] Furthermore, in step 2, the construction of the graph structure corresponding to the CSI fingerprint features based on node features, edge features, and adjacency matrix specifically involves:

[0038] Assuming the user terminal is currently located at a reference point, it receives a signal from N in an indoor positioning scenario. AP The transmitted signals from each access point correspond to N. AP One CSI fingerprint feature;

[0039] With N AP Each access point is a node, and the amplitude and phase features corresponding to the access point are used as node features. The edge weights are used as edge features. The adjacency relationship of the nodes is represented by the adjacency matrix A, thus obtaining the graph structure corresponding to all CSI fingerprint features at the current reference point.

[0040] Furthermore, in step 2, the image category label specifically refers to:

[0041] Y∈{0,1,...,N RP -1}

[0042] Where Y represents the graph category label, and each graph structure has one and only one graph category label. Assuming that the current user terminal is located at the k-th reference point, the graph category label corresponding to the graph structure at that reference point is Y = k-1.

[0043] Furthermore, in step 3, the output features of GNN1 and GNN2 are concatenated and input into the MLP. The MLP dynamically adjusts the weight parameters and outputs the prediction result, specifically as follows:

[0044] The output feature of GNN1 is vector r1, and the output feature of GNN2 is vector r2. Vector r1 and vector r2 correspond to the magnitude feature of input GNN1 and the phase feature of input GNN2, respectively. Vector r1 and vector r2 are concatenated and then input into MLP.

[0045] In indoor positioning scenarios, based on the different degrees to which amplitude and phase features are affected by noise, the MLP dynamically adjusts the weights of the vector r1 corresponding to the amplitude feature and the vector r2 corresponding to the phase feature in the prediction result. By assigning a larger weight to the vector corresponding to the feature less affected by noise, the final prediction result is output.

[0046] Secondly, in order to achieve the above-mentioned objectives, the present invention also provides a CSI indoor positioning device based on graph neural networks, including an indoor positioning scene construction module, a graph dataset acquisition module, a CSI-GNN algorithm model training module, and a CSI-GNN algorithm model actual inference module.

[0047] The indoor positioning scene construction module is used to construct an indoor positioning scene, including: arranging a number of access points for transmitting signals in the indoor positioning scene, setting a positioning area in the indoor positioning scene, and arranging a number of reference points for representing the location of the user terminal in the positioning area.

[0048] The graph dataset acquisition module is used by the user terminal to receive transmitted signals from several access points at different reference points and extract CSI fingerprint features respectively. The CSI fingerprint features are then dimensionality-reduced to obtain node features, which include amplitude features and phase features. Edge features and adjacency matrices are obtained based on the topological information between access points. The graph structure corresponding to the CSI fingerprint features is constructed based on the node features, edge features, and adjacency matrix, and graph category labels are labeled. The set of graph structures and their corresponding graph category labels is the graph dataset.

[0049] The CSI-GNN algorithm model training module is used to train the CSI-GNN algorithm model containing GNN1, GNN2 and MLP offline using a graph dataset. The process includes: inputting amplitude features into GNN1, inputting phase features into GNN2, inputting edge features and adjacency matrix into GNN1 and GNN2, concatenating the output features of GNN1 and GNN2 and inputting them into MLP, dynamically adjusting the weight parameters of MLP, outputting the prediction result, and minimizing the loss function between the prediction result and the graph category label, or reaching the maximum number of iterations, to obtain the trained CSI-GNN algorithm model.

[0050] The actual inference module of the CSI-GNN algorithm model is used to perform online positioning using the trained CSI-GNN algorithm model and output the real-time location of the user terminal.

[0051] Thirdly, to achieve the above-mentioned objectives, embodiments of the present invention also provide a CSI indoor positioning device based on a graph neural network, including a memory and a processor. The memory is used to store a computer program, and the processor is used to implement the CSI indoor positioning method based on a graph neural network provided in the first aspect of the present invention when the computer program is executed.

[0052] The beneficial effects of this invention are as follows:

[0053] (1) The CSI-GNN algorithm model proposed in this invention is based on CSI for indoor positioning. Compared with RSSI, CSI has many significant advantages. First, CSI provides more detailed information about the wireless channel, including the amplitude and phase changes of the signal. This information can more accurately depict the propagation path of the signal in complex indoor environments. For example, phase information can be used to identify and distinguish multipath propagation, which RSSI cannot achieve. This is because RSSI only provides an overall measure of signal strength and cannot reflect the interaction between the signal and the environment in detail. Second, the high-dimensionality of CSI makes it more effective in processing signals under multipath propagation and non-line-of-sight conditions. In addition, CSI provides more comprehensive information, which makes it perform better in indoor positioning systems based on machine learning and deep learning. These systems can use the detailed information of CSI to perform more complex feature extraction and environmental understanding.

[0054] (2) The CSI-GNN algorithm model proposed in this invention exhibits excellent adaptive capabilities. Specifically, when the amplitude or phase features are subject to significant noise interference, the CSI-GNN algorithm model dynamically adjusts the weight parameters of the MLP and prioritizes the use of features less affected by noise. This not only improves the accuracy and stability of the positioning system but also enhances the robustness of the CSI-GNN algorithm model in the face of variable environments and complex noise.

[0055] (3) The CSI-GNN algorithm addresses indoor positioning from a graph perspective, fully utilizing the topological information between access points. According to wireless communication theory, when the user terminal's location is fixed, the closer two access points are, the more similar their CSI (Communications in Memory) will be to the user terminal. This similarity information is transformed into edge weights in the graph. The graph convolutional layer aggregates node features using these edge weights, making adjacent nodes more similar in features. Compared to traditional CNN algorithms, CSI-GNN fully utilizes CSI information while avoiding the loss of topological information between access points, thus further improving the accuracy of indoor positioning. Attached Figure Description

[0056] Figure 1 This is a flowchart of the CSI indoor positioning method based on graph neural networks provided in an embodiment of the present invention.

[0057] Figure 2 This is a schematic diagram illustrating the specific process of the CSI indoor positioning method based on graph neural networks provided in this embodiment of the invention.

[0058] Figure 3 This is a schematic diagram of an indoor positioning scenario provided by an embodiment of the present invention.

[0059] Figure 4 This is a schematic diagram of the CSI-GNN algorithm provided in an embodiment of the present invention.

[0060] Figure 5 This is a schematic diagram of the GNN structure in the CSI-GNN algorithm provided in this embodiment of the invention.

[0061] Figure 6 This is a schematic diagram of the structure of the MLP in the CSI-GNN algorithm provided in this embodiment of the invention.

[0062] Figure 7 This is a schematic diagram of the structure of the CSI indoor positioning device based on graph neural network provided in an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0064] like Figure 1 , Figure 2 As shown, the embodiment provides a CSI indoor positioning method based on graph neural networks, including the following steps:

[0065] S110, Construct an indoor positioning scenario, including: setting up several access points for transmitting signals in the indoor positioning scenario, setting up a positioning area in the indoor positioning scenario, and setting up several reference points in the positioning area to represent the location of the user terminal.

[0066] First, an indoor positioning scenario is constructed. In this embodiment, a CSI-based indoor single-user positioning scenario is used as an example for illustration. Figure 3 As shown, N are distributed in an indoor single-user positioning scenario. AP There are N access points, and N are set within the location area. RP There are N reference points, and the physical locations of these reference points are known. A user is located at one of the reference points within the positioning area, and this user carries a terminal device (e.g., a mobile phone). The indoor positioning problem can be described as follows: based on the user's terminal device and N reference points, the user is positioned at one of the reference points. APThe CSI fingerprint characteristics between access points are used to determine which reference point the user is located at, and thus determine the user's physical location.

[0067] S120, the user terminal receives transmitted signals from several access points at different reference points and extracts CSI fingerprint features respectively. The CSI fingerprint features are dimensionality reduced to obtain node features, which include amplitude features and phase features. Edge features and adjacency matrix are obtained based on the topological information between access points. The graph structure corresponding to the CSI fingerprint features is constructed based on the node features, edge features and adjacency matrix and the graph category label is labeled. The set of graph structures and their corresponding graph category labels is the graph dataset.

[0068] When the user terminal is located at any reference point, the user terminal and N AP The CSI fingerprint characteristics between access points can be measured and expressed by the formula:

[0069]

[0070] Among them, CSI AP This indicates the value obtained from N at a certain reference point. AP A set of CSI fingerprint features for each access point, CSI j It is an N×M×K matrix representing the CSI fingerprint features between the user terminal and the j-th access point when the user is located at the current reference point, j = 1, 2, ..., N. AP N represents the number of receiving antennas for the user terminal, M represents the number of transmitting antennas for each access point, assuming that the number of transmitting antennas for all access points is the same, and K represents the number of subcarriers.

[0071] In indoor positioning scenarios, changing the user terminal's position from N RP CSI fingerprint features are collected at each reference point. Each CSI fingerprint feature at a reference point contains signal characteristics unique to that reference point, such as amplitude and phase information. The CSI fingerprint features collected at each reference point are then used to construct a fingerprint database.

[0072] For CSI fingerprint features in the fingerprint database, the amplitude and phase features corresponding to the transmitted signal received at a certain reference point from the access point are extracted as node features. Considering the topological information between access points, the physical distance between access points is converted into edge weights as edge features. Based on the node features, edge features, and adjacency matrix, a graph structure corresponding to the CSI fingerprint features at each reference point is constructed. In this embodiment, the graph structure specifically adopts a fully connected undirected graph, N... RP The set of fully connected undirected graphs corresponding to the CSI fingerprint features collected at each reference point constitutes the graph dataset. The construction of the fully connected undirected graphs is as follows:

[0073] With access points as nodes, the number of nodes is N. AP When the user terminal is located at a certain reference point, the CSI fingerprint feature between the j-th node and the user terminal is csi. j CSI j Dimensionality reduced to k-dimensional vector This results in two new vectors, la. j =[l j1 ,l j2 ,...,l jk ] and lp j =[θ j1 ,θ j2 ,...,θ jk ], where la j Corresponding CSI j Amplitude information, lp j Corresponding CSI j Phase information. la j and lp j The two types of node features that constitute the j-th node are: amplitude features and amplitude features. and phase characteristics

[0074] In indoor positioning scenarios, the locations of access points are fixed. The physical distance between access points is transformed into edge weights in a fully connected undirected graph, serving as edge features. The edge weight w between the i-th node and the j-th node is... ij It is expressed as follows:

[0075]

[0076] Where, d ij This represents the physical distance between the i-th access point and the j-th access point.

[0077] The adjacency matrix A of a fully connected undirected graph reflects the adjacency relationships between nodes. Based on these adjacency relationships and edge weights, the adjacency matrix A is constructed as follows:

[0078]

[0079] Where, when i≠j, a ij =w ij When i ≠ j, a ij =a ii =0.

[0080] After being represented by a graph structure, the indoor positioning problem is transformed into a graph classification problem: determining the graph's category based on its structural information (node ​​features, edge features, adjacency matrix). The number of categories is the same as the number of reference points, which is N. RPDefine the graph category label Y∈{0,1,...,N}. RP -1}, each fully connected undirected graph has one and only one graph category label. Assuming the user is located at the k-th reference point during sampling, the graph category label corresponding to the fully connected undirected graph at that reference point is Y = k-1.

[0081] S130. The CSI-GNN algorithm model, which includes GNN1, GNN2 and MLP, is trained offline using a graph dataset. This includes: inputting amplitude features into GNN1, inputting phase features into GNN2, inputting edge features and adjacency matrix into GNN1 and GNN2, concatenating the output features of GNN1 and GNN2 and inputting them into MLP, dynamically adjusting the weight parameters of MLP, outputting the prediction results, and minimizing the loss function between the prediction results and the graph category labels, or reaching the maximum number of iterations, to obtain the trained CSI-GNN algorithm model.

[0082] The structural block diagram of the CSI-GNN algorithm model designed in this invention is as follows: Figure 4 As shown. The indoor positioning method based on CSI-GNN is divided into offline and online stages. The graph dataset was constructed in S120, and then the CSI-GNN algorithm model was trained offline using the graph dataset. The relevant steps are described below:

[0083] 1): The amplitude feature LA and phase feature LP are used as the node features of GNN1 and GNN2 respectively, and the edge weights of the adjacency matrix A are used as the edge features of GNN1 and GNN2. The CSI-GNN algorithm model is then propagated forward.

[0084] GNN1 and GNN2 are GNNs with identical structures, as shown below. Figure 5 As shown. Each GNN contains three GraphConv graph convolutional layers. The input to each graph convolutional layer includes: node features, edge indices, and edge weights. The edge indices indicate the connections between nodes. The edge indices and edge weights are obtained from the adjacency matrix A. The graph convolutional layers update node features by aggregating the features of the node and its neighboring nodes. The aggregation method of the GraphConv graph convolutional layers is as follows:

[0085]

[0086] Where, x i Describe the node characteristics of node i, {x j |j∈N(i)} represents all neighboring nodes of node i. ji These are elements in the adjacency matrix A. W1 and W2 are the learnable parameters of the model. i ′ is the node feature of node i after being updated by the GraphConv graph convolutional layer.

[0087] Each GraphConv graph convolutional layer is followed by a BatchNorm layer. This layer is used to adjust and normalize the layer output, improving model stability and learning efficiency. A ReLU activation function is used after each BatchNorm layer to introduce non-linearity and enhance the model's ability to handle complex data. Finally, global average pooling is used to process the features of all nodes, summing them into a graph-level feature representation for graph classification tasks. The principle of global average pooling is as follows:

[0088]

[0089] The essence of global average pooling is to average the features of all nodes in each fully connected undirected graph to obtain a feature vector r that represents the graph-level features. The feature vector r is the output feature of the GNN.

[0090] 2): The output features of GNN1 and GNN2 are concatenated and then input into the MLP. The MLP contains three linear layers, with ReLU activation functions added between adjacent linear layers. The structure of the MLP is as follows: Figure 6 As shown. The output features of the MLP are denoted as Logits. Logits is an N RP The vector is dimensional, and the index of the maximum value in the vector is p. This indicates that the CSI-GNN algorithm model predicts that the user terminal is located at the reference point corresponding to the graph category label p.

[0091] 3): Calculate the cross-entropy loss between Logits and graph class label Y, perform backpropagation, calculate the gradient based on the cross-entropy loss, and update the learnable parameters of the model through the Adam optimizer.

[0092] 4) Repeat forward and backward propagation until the cross-entropy loss is less than the set threshold or the set number of iterations is reached to obtain the trained CSI-GNN algorithm model, which can be used for real-time indoor positioning prediction.

[0093] S140 uses a trained CSI-GNN algorithm model for online positioning and outputs the real-time location of the user terminal.

[0094] The trained CSI-GNN algorithm model is applied to localization tasks in indoor and arbitrarily complex scenarios. The proposed CSI-GNN algorithm model considers not only the amplitude and phase information of CSI but also the topological information of the access points. When the amplitude or phase features of CSI are significantly affected by noise, the method proposed in this invention can dynamically adjust the weight parameters of the MLP, prioritizing features less affected by noise, thereby improving the accuracy and robustness of the localization system. Therefore, this invention achieves good results not only in indoor localization scenarios but also in similar complex scenarios, ensuring excellent detection performance.

[0095] Based on the same inventive concept, embodiments of the present invention also provide a CSI indoor positioning device 700 based on a graph neural network, such as... Figure 7 As shown, it includes an indoor positioning scene construction module 710, a graph dataset acquisition module 720, a CSI-GNN algorithm model training module 730, and a CSI-GNN algorithm model actual inference module 740;

[0096] The indoor positioning scene construction module 710 is used to construct an indoor positioning scene, including: arranging several access points for transmitting signals in the indoor positioning scene, setting a positioning area in the indoor positioning scene, and arranging several reference points for representing the location of the user terminal in the positioning area.

[0097] The graph dataset acquisition module 720 is used for the user terminal to receive transmitted signals from several access points at different reference points and extract CSI fingerprint features respectively, perform dimensionality reduction processing on the CSI fingerprint features to obtain node features, the node features include amplitude features and phase features, obtain edge features and adjacency matrix according to the topological information between access points, construct the graph structure corresponding to the CSI fingerprint features according to the node features, edge features and adjacency matrix and label the graph category, the set of graph structure and its corresponding graph category label is the graph dataset;

[0098] The CSI-GNN algorithm model training module 730 is used to train the CSI-GNN algorithm model, which includes GNN1, GNN2 and MLP, offline using a graph dataset. The process includes: inputting amplitude features into GNN1, inputting phase features into GNN2, inputting edge features and adjacency matrix into GNN1 and GNN2, concatenating the output features of GNN1 and GNN2 and inputting them into MLP, dynamically adjusting the weight parameters of MLP, outputting the prediction results, and minimizing the loss function between the prediction results and the graph category labels, or reaching the maximum number of iterations, to obtain the trained CSI-GNN algorithm model.

[0099] The CSI-GNN algorithm model actual inference module 740 is used to perform online positioning using the trained CSI-GNN algorithm model and output the real-time location of the user terminal.

[0100] Based on the same inventive concept, the embodiment also provides a CSI indoor positioning device based on graph neural networks, including a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to implement the above-described CSI indoor positioning method based on graph neural networks when the computer program is executed.

[0101] It should be noted that the CSI indoor positioning device and CSI indoor positioning equipment based on graph neural networks provided in the above embodiments are based on the same concept as the CSI indoor positioning method embodiment based on graph neural networks. For details of their implementation process, please refer to the CSI indoor positioning method embodiment based on graph neural networks, which will not be repeated here.

[0102] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A CSI indoor positioning method based on graph neural networks, characterized in that, Includes the following steps: Step 1: Construct an indoor positioning scenario, including: setting up several access points for transmitting signals in the indoor positioning scenario, setting up a positioning area in the indoor positioning scenario, and setting up several reference points in the positioning area to represent the location of the user terminal. Step 2: The user terminal receives transmitted signals from several access points at different reference points and extracts CSI fingerprint features respectively. The CSI fingerprint features are dimensionality reduced to obtain node features, which include amplitude features and phase features. Edge features and adjacency matrix are obtained based on the topological information between access points. The graph structure corresponding to the CSI fingerprint features is constructed based on the node features, edge features and adjacency matrix and the graph category label is labeled. The set of graph structure and its corresponding graph category label is the graph dataset. Step 3: Train the CSI-GNN algorithm model offline using the graph dataset, which includes GNN1, GNN2 and MLP. This includes: inputting amplitude features into GNN1, phase features into GNN2, edge features and adjacency matrix into GNN1 and GNN2, concatenating the output features of GNN1 and GNN2 and inputting them into MLP, dynamically adjusting the weight parameters of MLP, outputting the prediction results, and minimizing the loss function between the prediction results and the graph category labels, or reaching the maximum number of iterations, to obtain the trained CSI-GNN algorithm model. The output features of GNN1 and GNN2 are concatenated and then input into the MLP. The MLP dynamically adjusts the weight parameters and outputs the prediction result, specifically: The output features of GNN1 are vectors r 1 The output features of GNN2 are vectors. r 2 ,vector r 1 sum vector r 2 The vectors correspond to the magnitude features of input GNN1 and the phase features of input GNN2, respectively. r 1 sum vector r 2 Input the concatenated data into MLP; In indoor positioning scenarios, based on the different degrees to which amplitude and phase features are affected by noise, the MLP dynamically adjusts the vector corresponding to the amplitude feature. r 1 Vectors corresponding to phase features r 2 The weights in the prediction results are assigned by giving larger weights to the vectors corresponding to features that are less affected by noise, and the final prediction results are output. Step 4: Use the trained CSI-GNN algorithm model to perform online localization and output the real-time location of the user terminal.

2. The CSI indoor positioning method based on graph neural networks according to claim 1, characterized in that, In step 2, the CSI fingerprint feature is expressed by the formula: in, CSI AP This indicates that the data obtained at a certain reference point comes from... N AP A set of CSI fingerprint features for each access point. csi j It is N × M × K The matrix represents the relationship between the user terminal at the current reference point and the first reference point. j CSI fingerprint characteristics between access points , N This indicates the number of receiving antennas on the user terminal. M This indicates the number of transmit antennas at each access point. K This indicates the number of subcarriers.

3. The CSI indoor positioning method based on graph neural networks according to claim 2, characterized in that, In step 2, the dimensionality reduction processing of the CSI fingerprint features is performed to obtain node features, which include amplitude features and phase features, specifically: right csi j Dimensionality reduction is performed to obtain k A dimensional vector, expressed by the formula: in, L j Indicates the first j The corresponding nodes k dimensional vector, l jt express L j No. t The amplitude of each element, θ jt express L j No. t The phase of each element; from k The amplitude and phase features are extracted from the dimensional vector, expressed by the formula: in, LA Indicates amplitude characteristics, LP Indicates phase characteristics, ,correspond csi j Amplitude information, ,correspond csi j Phase information.

4. The CSI indoor positioning method based on graph neural networks according to claim 2, characterized in that, In step 2, obtaining edge features based on the topological information between access points specifically involves: The topology information is based on the physical distance between access points; The edge weights are obtained based on the physical distance between access points, and can be expressed by the following formula: in, Represents edge weight, d ij Indicates the first i The first access point and the first j The physical distance between access points; Edge weights are used as edge features.

5. The CSI indoor positioning method based on graph neural networks according to claim 4, characterized in that, In step 2, based on the obtained edge weights, the topology information is used to construct an adjacency matrix using the adjacency relationships between access points, which is expressed by the formula: in, A Represents the adjacency matrix. a ij Indicates from the first i The node to the first j The edge features of each node, when i ≠ j hour, ,when i = j hour, .

6. The CSI indoor positioning method based on graph neural networks according to claim 3, 4, or 5, characterized in that, In step 2, the construction of the graph structure corresponding to the CSI fingerprint features based on node features, edge features, and adjacency matrix specifically involves: Assuming the user terminal is currently located at a reference point, it receives a signal from an indoor positioning scenario. N AP The transmitted signals from each access point correspond to... N AP One CSI fingerprint feature; by N AP Each access point is a node, with the amplitude and phase features corresponding to the access point as node features, and the edge weights as edge features. The adjacency relationship of the nodes is determined by the adjacency matrix. A The characterization process yields the graph structure corresponding to all CSI fingerprint features at the current reference point.

7. The CSI indoor positioning method based on graph neural networks according to claim 1, characterized in that, In step 2, the image category label specifically refers to: in, Y This represents the graph category label. Each graph structure has one and only one graph category label. Assuming the current user terminal is located at the [missing information], k If there are 1 reference point, then the graph category label corresponding to the graph structure at that reference point is: Y = k -1, set within the location area N RP A reference point.

8. A CSI indoor positioning device based on graph neural networks, characterized in that, It includes an indoor positioning scene construction module, a graph dataset acquisition module, a CSI-GNN algorithm model training module, and a CSI-GNN algorithm model actual inference module, to implement the CSI indoor positioning method based on graph neural network as described in any one of claims 1-7; The indoor positioning scene construction module is used to construct an indoor positioning scene, including: arranging a number of access points for transmitting signals in the indoor positioning scene, setting a positioning area in the indoor positioning scene, and arranging a number of reference points for representing the location of the user terminal in the positioning area. The graph dataset acquisition module is used by the user terminal to receive transmitted signals from several access points at different reference points and extract CSI fingerprint features respectively. The CSI fingerprint features are then dimensionality-reduced to obtain node features, which include amplitude features and phase features. Edge features and adjacency matrices are obtained based on the topological information between access points. The graph structure corresponding to the CSI fingerprint features is constructed based on the node features, edge features, and adjacency matrix, and graph category labels are labeled. The set of graph structures and their corresponding graph category labels is the graph dataset. The CSI-GNN algorithm model training module is used to train the CSI-GNN algorithm model containing GNN1, GNN2 and MLP offline using a graph dataset. The process includes: inputting amplitude features into GNN1, inputting phase features into GNN2, inputting edge features and adjacency matrix into GNN1 and GNN2, concatenating the output features of GNN1 and GNN2 and inputting them into MLP, dynamically adjusting the weight parameters of MLP, outputting the prediction result, and minimizing the loss function between the prediction result and the graph category label, or reaching the maximum number of iterations, to obtain the trained CSI-GNN algorithm model. The actual inference module of the CSI-GNN algorithm model is used to perform online positioning using the trained CSI-GNN algorithm model and output the real-time location of the user terminal.

9. A CSI indoor positioning device based on a graph neural network, comprising a memory and a processor, wherein the memory is used to store a computer program, characterized in that, The processor is configured to implement the CSI indoor positioning method based on graph neural networks as described in any one of claims 1-7 when executing the computer program.

Images