An indoor positioning method based on a graph neural network

Abstract

The application relates to an indoor positioning method based on a graph neural network, and relates to an indoor positioning method based on a graph neural network.The application aims to solve the problem that the positioning method based on a deep learning model in the prior art seldom considers the similarity and correlation between points in a fingerprint library, resulting in generally low positioning accuracy.The specific process of the indoor positioning method based on the graph neural network is as follows: step one, obtaining a training set;step two, constructing fingerprint library data based on the training set;step three, constructing edge features of the fingerprint library data based on the fingerprint library data to obtain complete graph data;step four, constructing a deep learning model;step five, obtaining a trained deep learning model based on the complete graph data;step six, obtaining a test set; and step seven, positioning the test set based on the trained deep learning model.The application is used in the field of indoor positioning based on a graph neural network.

Description

Technical Field

[0001] This invention relates to an indoor positioning method based on graph neural networks. Background Technology

[0002] Currently, outdoor positioning and navigation technologies based on the Global Positioning System (GPS) or cellular mobile networks are relatively mature. However, indoor environments often have numerous obstructions and obstacles, making satellite or cellular network signals vulnerable and thus preventing indoor positioning using GPS or cellular mobile network technologies. Since humans spend an average of 70% to 90% of their time indoors, we have a greater desire for positioning and tracking technologies in indoor environments. Indoor positioning primarily relies on technologies such as Wi-Fi, Bluetooth, visible light, and UWB. Bluetooth-based positioning technology is widely used due to its universal availability and low deployment costs. Some Bluetooth-based indoor positioning methods employ three-dimensional coordinates and triangulation algorithms based on received signal strength. However, due to the complex multipath effects and dynamic characteristics of wireless signal propagation in indoor environments, these ranging positioning methods based on received signal strength still struggle to achieve high positioning accuracy. To address the complex multipath effects and dynamic characteristics of wireless signal propagation in indoor environments, some researchers have adopted radio map matching methods (i.e., fingerprint-based positioning algorithms) to achieve better positioning accuracy than ranging positioning methods based on received signal strength. In recent years, deep learning has made tremendous progress and has been widely applied in fields such as computer vision and natural language processing. In computer vision, CNNs are used to automatically extract reasonable and rich features from raw images without the need for manually designed image feature extractors. In natural language processing, recurrent neural networks (RNNs) based on long short-term memory (LSTM) have attracted much attention due to their excellent performance in speech recognition systems, thanks to their ability to preserve temporal memory over long periods by using forgetting, input, and output gates. The excellent representational performance of deep learning technologies has led many experts and scholars to apply deep learning methods to indoor positioning. Some researchers have used classification and regression models based on deep neural networks such as CNN, LSTM, and ResNet. Although these deep learning models have achieved good positioning accuracy, they rarely consider the similarity and correlation between points in the fingerprint database, resulting in generally low positioning accuracy. In recent years, graph neural networks have made significant progress in areas such as social networks, knowledge graphs, and semantic segmentation because they consider the correlation between nodes in the graph. Therefore, this invention considers using graph neural network-based algorithms for indoor positioning. Summary of the Invention

[0003] The purpose of this invention is to address the problem that existing deep learning-based positioning methods rarely consider the similarity and correlation between points in the fingerprint database, resulting in generally low positioning accuracy. Therefore, this invention proposes an indoor positioning method based on graph neural networks.

[0004] The specific process of an indoor positioning method based on graph neural networks is as follows:

[0005] Step 1: Obtain the training set;

[0006] Step 2: Construct a fingerprint database based on the training set;

[0007] Step 3: Construct edge features of the fingerprint database based on the fingerprint database data to obtain complete graph data;

[0008] Step 4: Construct a deep learning model;

[0009] Step 5: Obtain the trained deep learning model based on the complete graph data;

[0010] Step 6: Obtain the test set;

[0011] Step 7: Locate the test set based on the trained deep learning model.

[0012] The beneficial effects of this invention are as follows:

[0013] 1. Establish a graph structure for Bluetooth strength data, taking into account the similarity and correlation between fingerprint database points;

[0014] 2. When constructing the graph data, the fingerprint database points are sparse. Gaussian interpolation was used to construct the fingerprint database for the points that were not collected in the middle.

[0015] 3. Modify the information transmission method of the graph attention network by adding edge information as a feature, which can improve the localization accuracy.

[0016] To address the indoor positioning problem, this invention proposes a positioning algorithm based on graph neural networks, such as... Figure 3 Compared with traditional positioning algorithms, the positioning algorithm of this invention improves positioning accuracy, achieving a positioning accuracy of 1.6 meters in a laboratory setting with densely arranged nodes, which is higher than the positioning accuracy of KNN and ResNet. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention;

[0018] Figure 2 A diagram of the deep learning model structure;

[0019] Figure 3 The graphs show the distribution functions obtained using different deep learning models. Detailed Implementation

[0020] Specific Implementation Method 1: The specific process of this indoor positioning method based on graph neural networks is as follows:

[0021] The purpose of this invention is to propose a deep learning method for indoor positioning based on Bluetooth, which extracts Bluetooth signal strength information for fingerprint database matching and positioning. Most fingerprint database positioning methods based on Bluetooth and WiFi do not consider the correlation between fingerprint database points that are close to each other, resulting in generally low positioning accuracy. Graph neural networks can solve the problem of correlation between nodes. By establishing a graph deep learning model, feature information from Bluetooth RSSI is extracted, thereby solving the problem of low indoor positioning accuracy.

[0022] The technical solution of this invention is as follows: To solve the above-mentioned technical difficulties in Bluetooth fingerprint database positioning, its overall flowchart is as follows. Figure 1 As shown.

[0023] The data collection app is developed for mobile devices, with underlying code in Java.

[0024] Step 1: Obtain the training set;

[0025] Step 2: Construct a fingerprint database based on the training set;

[0026] Step 3: Construct edge features of the fingerprint database based on the fingerprint database data to obtain complete graph data;

[0027] Step 4: Construct a deep learning model;

[0028] Step 5: Obtain the trained deep learning model based on the complete graph data;

[0029] Step 6: Obtain the test set;

[0030] Step 7: Locate the test set based on the trained deep learning model.

[0031] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the training set is obtained in step one; the specific process is as follows:

[0032] A fingerprint database region is constructed based on the positioning scenario (the positioning scenario is used as the fingerprint database region). A rectangular coordinate system XY is constructed in the fingerprint database region, with the X-axis coordinate value being x and the Y-axis coordinate value being y.

[0033] The fingerprint database area is M meters long and N meters wide, where M and N are integers (M = 15, N = 20).

[0034] The origin (x,y) is set at the bottom left corner (0,0) of the fingerprint database area.

[0035] In the fingerprint database area, (x+1,y), (x+2,y), (x+3,y), ..., (x+M,y) and (x,y+1), (x,y+1), (x,y+1), ..., (x,y+N) are taken as sampling points respectively;

[0036] Sampling points are defined as 1 meter apart in the fingerprint database area;

[0037] BLE Bluetooth beacons are deployed at each sampling point, and Bluetooth signal strength data is obtained based on the BLE Bluetooth beacons. Each sampling point is sampled for 30 seconds.

[0038] The other steps and parameters are the same as in Specific Implementation Method 1.

[0039] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that step two involves constructing a fingerprint database based on the training set; the specific process is as follows:

[0040] Each Bluetooth beacon acquires L Bluetooth signal strength data points, where 20 ≤ L ≤ 30;

[0041] The average value of L Bluetooth signal strength data is taken, and the average value is used as the standardized Bluetooth signal strength data obtained from each Bluetooth beacon, which is then used as an element in the fingerprint database.

[0042] Each sampling point contains standardized Bluetooth signal strength data obtained from j Bluetooth beacons, which is used as a row in the fingerprint database, where 30≤j≤50;

[0043] A single Bluetooth beacon acquires a large amount of signal strength data, and the average of these data points is used to obtain the RSSI signal strength data for a single Bluetooth beacon. 1,1 Each row in the matrix contains one sampling point;

[0044] Construct fingerprint database data as shown in Formula 1:

[0045]

[0046] Where RSSI represents fingerprint database data, rssi 1,1 This represents the standardized Bluetooth signal strength data obtained from the first Bluetooth beacon in the first sampling point, rssi. 1,j rssi represents the standardized Bluetooth signal strength data obtained from the j-th Bluetooth beacon in the first sampling point. i,1 rssi represents the standardized Bluetooth signal strength data obtained from the first Bluetooth beacon at the i-th sampling point. i,j This represents the standardized Bluetooth signal strength data obtained from the j-th Bluetooth beacon at the i-th sampling point, where j represents the number of Bluetooth beacons and i represents the total number of sampling points.

[0047] When the Bluetooth signal strength data of a certain sampling point is not available, Gaussian interpolation is used to interpolate the fingerprint database so that the distance between sampling points in the fingerprint database is 1 meter.

[0048] The formula for Gaussian interpolation is:

[0049] h a = L1(a)×h1+L2(a)×h2+…+L i (a)×h i

[0050]

[0051] Among them, h a h1 represents the a-th sampling point (the a-th row in the RSSI fingerprint database), h2 represents the 1st sampling point (the 1st row in the RSSI fingerprint database), and h3 represents the 2nd sampling point (the 2nd row in the RSSI fingerprint database). i Let L represent the i-th sampling point (the i-th row in the RSSI fingerprint database), L1(a) represent the standardized Bluetooth signal strength data in the first sampling point, and L2(a) represent the standardized Bluetooth signal strength data in the second sampling point. i (a) represents the standardized Bluetooth signal strength data at the i-th sampling point, L k (a) represents the standardized Bluetooth signal strength data in the kth sampling point, where k represents the kth sampling point, 1≤k≤i, and i represents the total number of sampling points.

[0052] In actual data collection, some experimental scenarios have sparse fingerprint database data due to object occlusion. Therefore, Gaussian interpolation is used to interpolate the sparse fingerprint database so that the distance between sampling points in the fingerprint database is about 1 meter.

[0053] Other steps and parameters are the same as in specific implementation method one or two.

[0054] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that, in step three, edge features of the fingerprint database are constructed based on the fingerprint database data to obtain complete graph data; the specific process is as follows:

[0055] Edge information is established for sampling points in the fingerprint database with a distance of 2 meters as a constraint (the sampling points in the fingerprint database are 1 meter apart, but here the edge information is established between sampling points with a distance of 2 meters between them). A grid (3×3 area) is formed based on the sampling points and edge information. Location labels are affixed to the sampling points that are 2 meters apart in the grid to obtain complete image data (the sampling points include Bluetooth signal strength data).

[0056] The establishment of edge features in the fingerprint database needs to meet certain constraints, and a certain distance is used as a constraint to establish the edge features of the data;

[0057] Since the data is mainly concentrated in the range of -70dBm to -90dBm, the fingerprint database establishes edge information with a distance of 2 meters as a constraint, based on the Bluetooth RSSI propagation model. This ensures that the RSSI variation remains between 3 and 4dBm. If the constraint distance is too large, the RSSI variation will be too large, making it difficult to extract features. If the constraint distance is too small, the edge features will be too few, and the graph neural network will not be able to achieve good training results.

[0058] Nine adjacent fingerprint database sampling points are used to form a grid (3×3 area), and labels are attached as the input of the label network. For adjacent classification regions, we reuse sampling points (one of the sampling points can be used as the adjacent sampling point of the previous sampling point or the adjacent sampling point of the next sampling point), which can prevent the loss of the localization area.

[0059] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0060] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that step four involves constructing a deep learning model; the specific process is as follows:

[0061] The deep learning model consists of, in sequence, a GAT graph attention network, a ReLU activation function layer, a GAT graph attention network, a ReLU activation function layer, a Dropout layer, a GAT graph attention network, and a Softmax layer.

[0062] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0063] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that step five involves obtaining a trained deep learning model based on complete graph data; the specific process is as follows:

[0064] The complete graph data obtained in step three is used as the input to the deep learning model constructed in step four, and the center point position of the complete graph data is used as the output of the deep learning model to obtain the trained deep learning model.

[0065] There will be a non-linear relationship between the complete map data and the specific location;

[0066] Network structure and optimization:

[0067] The GAT (Graph Attention Network) is employed; compared to GCN (Graph Concentration Network), GAT is better suited for localization scenarios. While GCN directly averages the information between neighboring nodes during transmission, localization is problematic because the correlation between closer nodes is much greater than that between farther nodes, and the importance of neighboring nodes to the current node may differ. Using node averaging would introduce significant errors, hence the introduction of a graph attention mechanism. This mechanism quantifies the influence of neighboring nodes on the current node as an attention coefficient. The core difference between GAT and GCN lies in how they collect and aggregate feature representations of neighboring nodes at a distance of 1. Essentially, GAT simply replaces the normalization function of the original GCN with a neighboring node feature aggregation function using attention weights. The input to each attention layer is the node's own features (i.e., the constructed feature image).

[0068] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0069] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that step six involves obtaining the test set; the specific process is as follows:

[0070] Step 61: Pedestrians walk at a constant speed within the fingerprint database area while holding their mobile phones;

[0071] For each timestamp (sampling frequency, e.g., 5s), j′ Bluetooth signals are obtained after standardization, 20≤j′≤30 (if a mobile phone contains many Bluetooth devices, e.g., 10, then the Bluetooth signal strength data of 10 Bluetooth devices must be obtained).

[0072] If a Bluetooth device does not obtain standardized Bluetooth signal strength data, then there is a loss of Bluetooth signal strength data (the phone contains many Bluetooth devices, such as 10, but only receives the signal strength data of 8 Bluetooth devices). Select the Bluetooth signal strength data that is 5 to 10 dB smaller than the smallest Bluetooth signal strength data in the training set as the lost Bluetooth signal strength data.

[0073] The process by which each Bluetooth device acquires standardized Bluetooth signal strength data is as follows:

[0074] For each Bluetooth device, L′ Bluetooth signal strength data are acquired. The average value of the L′ Bluetooth signal strength data is taken as the normalized Bluetooth signal strength data acquired for each Bluetooth device, where 30≤L′≤50.

[0075] Step 6.2: One timestamp corresponds to one sampling point. Connect two adjacent sampling points to form an edge feature. The sampling points and edge features form a grid (e.g., a 3×3 region) to obtain graph data (one graph).

[0076] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0077] Specific Implementation Method Eight: This implementation method differs from any of Specific Implementation Methods One through Seven in that step Seven involves locating the test set based on the trained deep learning model; the specific process is as follows:

[0078] Input the graph data obtained in step six into the trained deep learning model, and the trained deep learning model outputs the coordinates of the center point of the graph data.

[0079] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0080] This invention consists of two stages: offline training and online prediction;

[0081] Offline system training converts Bluetooth features with bit tags into the aforementioned graph data, which serves as the input to the graph neural network classification model. After offline network training, a non-linear relationship will emerge between the Bluetooth feature images and specific locations.

[0082] In the online prediction phase, the model's network weights have been trained; Bluetooth feature images without location labels are used as input to the Bluetooth classification model; the probability of each location represents the similarity between the sample and all possible measurement points. The final estimated location is calculated based on the weights of the output class location labels.

[0083] The overall network structure is as follows Figure 2 As shown.

[0084] Example:

[0085] Bluetooth beacons were deployed in laboratory and underground parking garage scenarios, using BLE Bluetooth beacons. Compared to regular Bluetooth, BLE Bluetooth Low Energy offers advantages such as faster connection speeds and lower power consumption. A three-layer GAT network structure was employed, using the Adam optimizer, with the loss function set to NLLLoss and the learning rate set to 0.001, for 4000 training rounds.

[0086] like Figure 3 Compared with traditional positioning algorithms, the positioning algorithm of this invention improves positioning accuracy, achieving a positioning accuracy of 1.6 meters in a laboratory setting with densely arranged nodes, which is higher than the positioning accuracy of KNN and ResNet.

[0087] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. An indoor positioning method based on graph neural networks, characterized in that: The specific process of the method is as follows: Step 1: Obtain the training set; Step 2: Construct a fingerprint database based on the training set; the specific steps are as follows: Acquiring Bluetooth signal data via Bluetooth beacons; the process is as follows: When the Bluetooth signal strength data of a certain sampling point is not available, Gaussian interpolation is used to interpolate the fingerprint database so that the distance between sampling points in the fingerprint database is 1 meter. The formula for Gaussian interpolation is: in, Indicates the first One sampling point, This indicates the first sampling point. This indicates the second sampling point. Indicates the first One sampling point, This represents the standardized Bluetooth signal strength data at the first sampling point. This represents the standardized Bluetooth signal strength data at the second sampling point. Indicates the first Standardized Bluetooth signal strength data from each sampling point Indicates the first Standardized Bluetooth signal strength data from each sampling point Indicates the first One sampling point, , Represents the total number of sampling points; Step 3: Construct edge features from the fingerprint database to obtain complete graph data; the specific process is as follows: Edge information is established by constraining the sampling points in the fingerprint database to a distance of 2 meters. A grid is formed based on the sampling points and edge information. Location labels are affixed to adjacent sampling points within the grid that are 2 meters apart to obtain complete image data. Step 4: Construct a deep learning model; the specific process is as follows: The deep learning model consists of, in sequence, a GAT graph attention network, a ReLU activation function layer, a GAT graph attention network, a ReLU activation function layer, a Dropout layer, a GAT graph attention network, and a Softmax layer; Step 5: Obtain the trained deep learning model based on the complete graph data; Step 6: Obtain the test set; Step 7: Locate the test set based on the trained deep learning model.

2. The indoor positioning method based on graph neural networks according to claim 1, characterized in that: The training set is obtained in step one; the specific process is as follows: A fingerprint database region is constructed based on the localization scenario. A Cartesian coordinate system XY is then built within the fingerprint database region, with the X-axis coordinate value being... The Y-axis coordinate is ; The fingerprint database area is M meters long and N meters wide, where M and N are integers. The lower left corner of the fingerprint database area As the origin ; Take fingerprints from the fingerprint database area respectively. and As sampling points; BLE Bluetooth beacons are deployed at each sampling point, and Bluetooth signal strength data is obtained based on the BLE Bluetooth beacons. Each sampling point is sampled for 30 seconds.

3. The indoor positioning method based on graph neural networks according to claim 2, characterized in that: Step two involves constructing a fingerprint database based on the training set; the specific process is as follows: Each Bluetooth beacon acquires Bluetooth signal strength data, ; right The average value of each Bluetooth signal strength data point is used as the standardized Bluetooth signal strength data obtained from each Bluetooth beacon, and this average value is used as an element in the fingerprint database. Each sampling point contains The standardized Bluetooth signal strength data obtained from each Bluetooth beacon is used as a row in the fingerprint database. ; Construct fingerprint database data as shown in Formula 1: RSSI represents fingerprint database data. This represents the standardized Bluetooth signal strength data obtained from the first Bluetooth beacon in the first sampling point. Indicates the first sampling point. Standardized Bluetooth signal strength data obtained from Bluetooth beacons Indicates the first The standardized Bluetooth signal strength data obtained from the first Bluetooth beacon in each sampling point. Indicates the first The sampling point of the nth sampling point Standardized Bluetooth signal strength data obtained from Bluetooth beacons Represents the number of Bluetooth devices. Represents the total number of sampling points; When the Bluetooth signal strength data of a certain sampling point is not available, Gaussian interpolation is used to interpolate the fingerprint database so that the distance between sampling points in the fingerprint database is 1 meter. The formula for Gaussian interpolation is: in, Indicates the first One sampling point, This indicates the first sampling point. This indicates the second sampling point. Indicates the first One sampling point, This represents the standardized Bluetooth signal strength data at the first sampling point. This represents the standardized Bluetooth signal strength data at the second sampling point. Indicates the first Standardized Bluetooth signal strength data from each sampling point Indicates the first Standardized Bluetooth signal strength data from each sampling point Indicates the first One sampling point, , This represents the total number of sampling points.

4. The indoor positioning method based on graph neural networks according to claim 3, characterized in that: In step five, a trained deep learning model is obtained based on complete graph data; the specific process is as follows: The complete graph data obtained in step three is used as the input to the deep learning model constructed in step four, and the center point position of the complete graph data is used as the output of the deep learning model to obtain the trained deep learning model.

5. The indoor positioning method based on graph neural networks according to claim 4, characterized in that: Step six involves obtaining the test set; the specific process is as follows: Step 61: Pedestrians walk at a constant speed within the fingerprint database area while holding their mobile phones; Each timestamp Each Bluetooth device obtains standardized Bluetooth signal strength data. ; If a Bluetooth signal strength indicator (SSI) is not available after standardization, then Bluetooth SSI data loss is considered. The indicator selected should be smaller than the minimum SSI in the training set. dB Bluetooth signal strength data, as lost Bluetooth signal strength data; The process by which each Bluetooth device acquires standardized Bluetooth signal strength data is as follows: Each Bluetooth acquisition Bluetooth signal strength data, for The average value of each Bluetooth signal strength data point is used as the standardized Bluetooth signal strength data for each Bluetooth device. ; Step 6.2: One timestamp corresponds to one sampling point. Connect two adjacent sampling points to form an edge feature. The sampling points and edge features form a grid to obtain graph data.

6. The indoor positioning method based on graph neural networks according to claim 5, characterized in that: In step seven, the test set is located based on the trained deep learning model; the specific process is as follows: Input the graph data obtained in step six into the trained deep learning model, and the trained deep learning model outputs the coordinates of the center point of the graph data.

Images