Wireless positioning method based on residual image attention neural network

Abstract

The invention relates to the technical field of wireless fingerprint positioning, in particular to a wireless positioning method based on a residual image attention neural network, which comprises the following steps of: acquiring fingerprint data; image signal processing; designing a residual image neural network; using a Softmax classifier to perform position prediction according to an embedded vector generated by the graph network; the method has good performance on fingerprint map construction and positioning tasks in a complex environment, high-precision position estimation in the complex environment is realized, and high-quality position service can be provided; and meanwhile, the capability of deep fingerprint positioning to adapt to environment change is improved, and the method has good positioning performance in an indoor complex and changeable dynamic scene.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless fingerprint positioning, and in particular to a wireless positioning method based on a residual graph attention neural network. Background Art

[0002] High-precision and highly reliable indoor and outdoor positioning systems have become an urgent need for national security, economic construction, and social development. Therefore, in order to meet the needs of high-quality location services, it is necessary to mine spatiotemporal feature information in the massive data of the Internet of Everything for location calculation. Currently, fingerprint positioning based on deep learning is one of the key technologies with the greatest potential for large-scale applications. By designing deep neural networks (DNN) to extract fingerprint features and build wireless maps, the intrinsic relationship between signal features and location information can be better revealed, effectively improving the representation ability of fingerprint features and positioning accuracy.

[0003] The existing fingerprint feature recognition positioning method is not mature in commercial applications. In a complex and changeable positioning environment, wireless signal propagation is easily affected by external factors, which has a serious hindering effect on the efficient extraction of universal, robust and accurate fingerprint features in wireless fingerprint positioning. In addition, the establishment of fingerprint maps for different scenarios usually takes a lot of time. Since the existing algorithms are sensitive to changes in spatiotemporal characteristics, slight changes in time and environment will cause the fingerprint map to fail, and further increase the cost of fingerprint map establishment and maintenance; this greatly limits its application in actual systems.

[0004] Therefore, in view of the fact that the above-mentioned existing algorithms are sensitive to changes in spatiotemporal characteristics, and slight changes in time and environment will cause the fingerprint map to fail and further increase the cost of establishing and maintaining the fingerprint map, a wireless positioning method based on the residual graph attention neural network can be designed. Summary of the invention

[0005] In order to overcome the problem that existing algorithms are sensitive to changes in spatiotemporal features, slight changes in time and environment will cause the fingerprint map to fail and further increase the cost of establishing and maintaining the fingerprint map.

[0006] The technical solution of the present invention is: a wireless positioning method based on residual graph attention neural network, comprising the following steps: S1: fingerprint data collection: gridding the positioning area, setting reference points (RP), collecting physical layer channel state information (CSI) at each sampling position, collecting radio frequency fingerprint data, and establishing a data set of position-fingerprint data relationship where x i is the magnitude vector of multi-antenna multi-subcarrier CSI, yi is its corresponding position label, and N represents the number of all data in the dataset.

[0007] S2: Graph signal processing: The CSI fingerprint data is constructed into a graph data format, and the graph data processing method is used to obtain the CSI graph signal data G as the input of the deep graph model to learn the graph embedding feature representation;

[0008] S2.1: Take each CSI magnitude vector x as a node and build an unweighted graph is the set of all nodes in the graph, ε is the set of all edges in the graph, and any node v i The set of neighboring nodes is e ij ∈ε represents the node v i With any adjacent node v j The edge between

[0009] S2.2: Calculate the degree d of the node ij =∑ j A ij , thus obtaining the degree matrix D, where the elements d on the diagonal of the degree matrix ij =∑ j A ij , the rest of the elements are 0;

[0010] S2.3: Calculate the graph Laplacian matrix L = DA from the degree matrix D and the adjacency matrix A, and use the normalized Laplacian matrix Perform graph processing on CSI fingerprint data to obtain CSI graph structure data As input to the deep model g(G, A);

[0011] S3: Residual Graph Neural Network Design: Using the residual graph attention network, we design a multi-layer multi-head attention mechanism to aggregate neighborhood universal information and learn the graph embedding representation of CSI fingerprint data;

[0012] S4: Position estimation: Use the Softmax classifier to predict the position based on the embedding vector generated by the graph network and output the classification probability; its mathematical expression is as follows:

[0013] z'=σ(Wz 0 +b)

[0014]

[0015] Among them, z 0is the embedding vector output by the residual graph attention network, W and b represent a set of parameters that can be updated by the machine learning algorithm, σ(·) represents the ReLU function, and P(y=m|x) represents the probability that a certain position belongs to category m given a CSI signal input x.

[0016] Preferably, the adjacency matrix A represents the connection relationship between nodes, wherein the value of an element is 1, indicating that the corresponding two nodes are connected, otherwise it is 0, indicating that they are not connected.

[0017] Preferably, the residual graph attention network architecture consists of several residual graph attention network modules, each of which contains a multi-layer perceptron and a graph neural network structure.

[0018] As a preferred option, the residual graph attention network constructs a K-head attention mechanism. Through the attention mechanism, according to the input data of the module and A, estimate the similarity of adjacent nodes and obtain the weight of graph network aggregation, that is, any adjacent node v i and v j The attention coefficient α ij , where h i Represents node v i The node feature vector of is the set of real numbers, Indicates h i is an F-dimensional vector.

[0019] As a preferred method, the attention mechanism calculation method is

[0020]

[0021] Where W and a represent a set of parameters that can be updated by the machine learning algorithm, a T represents the transpose of matrix a, Represents node v i All adjacent nodes of . Thus, we can get any adjacent node v i and v j The attention coefficient α ij .

[0022] Preferably, according to K groups of attention coefficients, K groups of GNN networks are used to process and aggregate the feature vectors of the node and its adjacent nodes, and then the outputs of the K-head attention graph network are aggregated through a pooling step; finally, the input is dimensionally changed through a fully connected layer and added to the output of the K-head attention graph network to form a residual structure.

[0023] As a preference, the non-last layer network uses concatenated pooling, expressed as:

[0024]

[0025] and Represent the input signal of node i at the lth layer and the l+1th layer respectively, and They represent a set of parameters that can be updated by the machine learning algorithm, Indicates concatenating and pooling the output of the multi-head network. Represents the ELU activation function.

[0026] As a preference, the last layer of the network uses average pooling, which is expressed as:

[0027]

[0028] Indicates that the output of the multi-head network is average pooled.

[0029] The beneficial effects of the present invention are as follows: it has good performance in fingerprint map construction and positioning tasks in complex environments, realizes high-precision position estimation in complex environments, and can provide high-quality location services; at the same time, it improves the ability of deep fingerprint positioning to adapt to environmental changes, shows a certain robustness to changes in spatiotemporal characteristics, and exhibits good positioning performance for changing scenes; there is no need to repeatedly collect feature maps, which greatly reduces the cost of maintaining feature maps. DETAILED DESCRIPTION

[0030] The present invention will be further described below in conjunction with the embodiments.

[0031] The present invention provides an embodiment: a wireless positioning method based on a residual graph attention neural network, comprising the following steps: fingerprint data collection: gridding the positioning area, setting reference points (RP), collecting physical layer channel state information (CSI) at each sampling position, collecting radio frequency fingerprint data, and establishing a data set of position-fingerprint data relationship Graph signal processing: Use graph data processing methods to obtain CSI graph signal data G as the input of the deep graph model to learn graph embedding feature representation; use each CSI amplitude vector x as a node to establish an unweighted graph Calculate the degree d of the node ij =∑ j A ij , thus obtaining the degree matrix D, where the elements d on the diagonal of the degree matrix ij =∑ j A ij , the remaining elements are 0; the graph Laplacian matrix L = DA is calculated by the degree matrix D and the adjacency matrix A, and the normalized Laplacian matrix is ​​used Perform graph processing on CSI fingerprint data to obtain CSI graph structure data As the input of the deep model g(G, A); Residual graph neural network design: adopt the residual graph attention network and design the K-head attention mechanism. Through the attention mechanism, according to the input data of the module and A, estimate the similarity of adjacent nodes and obtain the weight of graph network aggregation; the calculation method is:

[0032]

[0033] According to K groups of attention coefficients, K groups of GNN networks are used to process and aggregate the feature vectors of nodes and their adjacent nodes, and then the output of the K-head attention graph network is aggregated through the pooling step; finally, the input is changed in dimension through the fully connected layer and added to the output of the K-head attention graph network to form a residual structure; the non-last layer network uses concatenated pooling, which is expressed as:

[0034]

[0035] The last layer of the network uses average pooling, which is expressed as:

[0036]

Claims

1. A wireless positioning method based on a residual graph attention neural network includes the following steps: S1: Fingerprint data collection: Divide the positioning area into grids, set reference points (RP), collect physical layer channel state information (CSI) at each sampling location, collect RF fingerprint data, and establish a data set of location-fingerprint data relationship where x i is the magnitude vector of multi-antenna multi-subcarrier CSI, y i is its corresponding position label, and N represents the number of all data in the data set; S2: Graph signal processing: The CSI fingerprint data is constructed into a graph data format, and the graph data processing method is used to obtain the CSI graph signal data G as the input of the deep graph model to learn the graph embedding feature representation; S2.1: Take each CSI magnitude vector x as a node and build an unweighted graph is the set of all nodes in the graph, ε is the set of all edges in the graph, and any node v i The set of neighboring nodes is e ij ∈ε represents the node v i With any adjacent node v j The edge between S2.2: Calculate the degree d of the node ij =∑ j A ij , thus obtaining the degree matrix D, where the elements d on the diagonal of the degree matrix ij =∑ j A ij , the rest of the elements are 0; S2.3: Calculate the graph Laplacian matrix L = DA from the degree matrix D and the adjacency matrix A, and use the normalized Laplacian matrix Perform graph processing on CSI fingerprint data to obtain CSI graph structure data As input to the deep model g(G, A); S3: Residual Graph Neural Network Design: Using the residual graph attention network, we design a multi-layer multi-head attention mechanism to aggregate neighborhood universal information and learn the graph embedding representation of CSI fingerprint data; S4: Position estimation: Use the Softmax classifier to predict the position based on the embedding vector generated by the graph network and output the classification probability; its mathematical expression is as follows: z'=σ(Wz 0 +b) Among them, z 0 is the embedding vector output by the residual graph attention network, W and b represent a set of parameters that can be updated by the machine learning algorithm, σ(·) represents the ReLU function, and P(y=m|x) represents the probability that a certain position belongs to category m given a CSI signal input x.

2. The wireless positioning method based on residual graph attention neural network according to claim 1 is characterized in that: The adjacency matrix A represents the connection relationship between nodes, where the value of an element is 1, indicating that the corresponding two nodes are connected, otherwise it is 0, indicating that they are not connected.

3. The wireless positioning method based on residual graph attention neural network according to claim 1, characterized in that: The residual graph attention network architecture consists of several residual graph attention network modules, each of which contains a multi-layer perceptron and graph neural network structure.

4. The wireless positioning method based on residual graph attention neural network according to claim 1, characterized in that: The residual graph attention network constructs a K-head attention mechanism. Through the attention mechanism, according to the input data H = {h1, h2, ... h N }, and A, estimate the similarity of adjacent nodes and obtain the weight of graph network aggregation, that is, any adjacent node v i and v j The attention coefficient α ij , where h i Represents node v i The node feature vector of is the set of real numbers, Indicates h i is an F-dimensional vector.

5. The wireless positioning method based on residual graph attention neural network according to claim 4 is characterized in that: The attention mechanism calculation method is Where W and a represent a set of parameters that can be updated by the machine learning algorithm, a T represents the transpose of matrix a, Represents node v i All adjacent nodes of . Thus, we can get any adjacent node v i and v j The attention coefficient α ij .

6. The wireless positioning method based on residual graph attention neural network according to claim 1, characterized in that: According to K groups of attention coefficients, K groups of GNN networks are used to process and aggregate the feature vectors of the node and its adjacent nodes, and then the output of the K-head attention graph network is aggregated through the pooling step; finally, the input is dimensionally changed through the fully connected layer and added to the output of the K-head attention graph network to form a residual structure.

7. The wireless positioning method based on residual graph attention neural network according to claim 1, characterized in that: The non-last layer network uses concatenated pooling, and the expression is: and Represent the input signal of node i at the lth layer and the l+1th layer respectively, and They represent a set of parameters that can be updated by the machine learning algorithm, Indicates concatenating and pooling the output of the multi-head network. Represents the ELU activation function.

8. The wireless positioning method based on residual graph attention neural network according to claim 1, characterized in that: The last layer of the network uses average pooling, which is expressed as: Indicates that the output of the multi-head network is average pooled.

Images