A visible light positioning method and device based on graph neural network

Abstract

The present invention discloses a visible light positioning method and device based on a graph neural network. The method comprises: constructing a visible light positioning model to be trained through a long short-term memory neural network on the basis of a graph neural network guided by an attention mechanism; inputting the collected real-time received signal strength values into the trained visible light positioning model to obtain the real-time position information of the object to be positioned. The graph neural network adopted by the present invention can simultaneously consider multiple positioning lights and their topological structures to optimize the signal, thereby suppressing the existing thermal noise, shot noise and ambient light noise; adopting an attention mechanism to guide the representation learning of the graph neural network, so that the network pays more attention to high-precision positioning lights, thereby improving the accuracy and robustness of positioning; modeling the time dependency between adjacent points in the trajectory through an LSTM network, so that the model can consider previous information when predicting the current position, further improving the accuracy and robustness of positioning.

Description

Technical Field

[0001] Embodiments of the present invention relate to the field of visible light positioning technology, and in particular to a visible light positioning method and device based on graph neural network. Background Art

[0002] Visible light positioning systems, as a good indoor positioning solution, have received widespread attention in academia and industry due to their advantages such as being green and environmentally friendly, high precision, strong resistance to electromagnetic interference, wide infrastructure coverage, and low cost.

[0003] Existing visible light positioning systems can be divided into camera-based visible light positioning systems and photodiode-based visible light positioning systems according to the type of receiver. The visible light positioning system based on photodiode (PD) is more popular because of its simple technology and low hardware cost. Among them, the visible light positioning system based on photodiode includes positioning methods based on three sides, based on arrival angle, and based on arrival time. Among them, the positioning method based on arrival time is difficult to achieve time synchronization, and the positioning method based on arrival angle generally requires multiple photodiodes, which increases the cost and difficulty of hardware design. The method based on three sides first uses Lambert's emission law to convert the received signal strength (RSS) value into the distance from the photodiode to the corresponding positioning light, and then estimates the position of the photodiode through the least squares principle. It is simple to implement and can achieve high accuracy. For details, see Figure 1 , Figure 1 This is a structural diagram of the visible light positioning system based on photodiode (PD).

[0004] However, visible light positioning systems based on photodiodes are often interfered with by various noises, the main ones of which include ambient light noise, shot noise, and thermal noise. The presence of these noises seriously hinders the positioning accuracy and robustness of existing visible light positioning systems. Ambient light noise is caused by the change of ambient light intensity over time and space. When the ambient light changes over time, resulting in a large difference in ambient light intensity between the calibration scene of the Lambertian parameters and the positioning scene, the Lambertian parameters obtained in the calibration scene will no longer be applicable to the positioning scene, resulting in a large positioning error; the impact of ambient light changes with space on the visible light positioning system is that the uneven distribution of ambient light intensity over space causes the RSS-distance relationship actually measured in the calibration scene to deviate from the ideal curve, thereby damaging the calibration of the Lambertian parameters and affecting the subsequent performance of the visible light positioning system.

[0005] Although modeling the ambient light by considering each lamp separately has a certain suppressive effect, it has huge limitations. For example, the existing solution integrates visible light positioning by collecting a large number of RSS values of the signal reception intensity under different ambient light intensities to fit the relationship between ambient light intensity and RSS. The RSS value received in the positioning scene is then mapped to the corresponding RSS value in the calibration scene through the fitted curve to suppress the influence of ambient light. However, this method requires the collection of a large amount of data in order to accurately fit the curve between RSS and ambient light, and the method is easily affected by data noise. In addition, shot noise and thermal noise are often modeled as additive Gaussian noise using random processes, making it difficult to estimate their specific values in each positioning. Therefore, the performance of existing visible light positioning systems will be affected by these noises. Summary of the Invention

[0006] The present invention provides a visible light positioning method and device based on graph neural network to improve the robustness and positioning accuracy of visible light positioning.

[0007] In a first aspect, the present invention provides a visible light positioning method based on a graph neural network, comprising:

[0008] S110. Based on the graph neural network guided by the attention mechanism, a visible light positioning model to be trained is constructed through a long short-term memory neural network.

[0009] S120: training the visible light positioning model to be trained to obtain a trained visible light positioning model;

[0010] S130: Input the collected real-time received signal strength values into the trained visible light positioning model to obtain real-time position information of the object to be positioned.

[0011] Optionally, the S110 includes:

[0012] Based on the collected positioning training data, the positioning training data is optimized using the spatial modeling capability of the graph neural network;

[0013] The attention mechanism is used to assign different weights to the positioning lights, thereby guiding the representation learning of the graph neural network;

[0014] Based on the Long Short-Term Memory (LSTM) neural network, the temporal relationship between the received signal strength values at adjacent positions on the same trajectory is modeled to obtain the visible light positioning model to be trained.

[0015] Optionally, based on the collected positioning training data, the positioning training data is optimized using the spatial modeling capability of a graph neural network, including:

[0016] Dynamically collect the received signal strength value of the positioning light and its corresponding position information through the signal acquisition device;

[0017] Constructing the collected received signal strength values with location tags into a graph structure;

[0018] Define the graph neural network structure;

[0019] The graph data structure is input into the graph neural network structure for processing, and the received signal strength value is optimized.

[0020] Optionally, the graph neural network structure includes 2 layers of GraphSage, each hidden layer has 16 hidden neurons, and each layer of the network includes a mean aggregator.

[0021] Optionally, the attention mechanism can be used to assign different weights to the location lights to guide the representation learning of the graph neural network, including:

[0022] Determine the signal-to-noise ratio corresponding to each positioning light based on the received signal strength value and noise corresponding to each positioning light;

[0023] Build an attention network;

[0024] Inputting the signal-to-noise ratio into the attention network to obtain the weight corresponding to each positioning light;

[0025] A weighted operation is performed on the weight and the received signal strength value corresponding to each positioning light.

[0026] Optionally, the attention network consists of two fully connected layers and one SoftMax layer.

[0027] Optionally, a long short-term memory (LSTM) neural network is used to model the temporal relationship between received signal strength values at adjacent locations on the same trajectory, including:

[0028] Collect weighted received signal strength values over a period of time;

[0029] Build LSTM network structure;

[0030] The weighted received signal strength value is fed into the LSTM network structure, and the output state of the last LSTM cell is retained.

[0031] Optionally, the LSTM network structure consists of a forget gate, an input gate, and an output gate.

[0032] In a second aspect, an embodiment of the present invention further provides a visible light positioning device based on a graph neural network, comprising:

[0033] A model building module, which is used to build a visible light positioning model to be trained using a long short-term memory neural network based on the graph neural network guided by the attention mechanism;

[0034] A model training module, used to train the visible light positioning model to be trained to obtain a trained visible light positioning model;

[0035] The real-time positioning module is used to input the collected real-time received signal strength values into the trained visible light positioning model to obtain the real-time position information of the object to be positioned.

[0036] Optionally, the model building module is specifically used to:

[0037] Based on the collected positioning training data, the positioning training data is optimized using the spatial modeling capability of the graph neural network;

[0038] The attention mechanism is used to assign different weights to the positioning lights, thereby guiding the representation learning of the graph neural network;

[0039] Based on the long short-term memory neural network (LSTM), the temporal relationship between the received signal strength values at adjacent positions on the same trajectory is modeled to obtain the visible light positioning model to be trained.

[0040] Beneficial effects of the present invention:

[0041] 1. The present invention uses graph neural networks for signal optimization and noise suppression in the visible light positioning system. Compared with the traditional method of only modeling noise in each positioning light separately, the present invention uses graph neural networks to simultaneously consider multiple positioning lights to optimize the RSS value and suppress noise. It can not only effectively alleviate the impact of uneven distribution of ambient light over time on the visible light positioning system, but also suppress the damage of uneven distribution of ambient light in space to the visible light positioning system.

[0042] 2. The present invention uses the attention mechanism to assign different weights to positioning lights according to their ranging accuracy, thereby guiding the graph neural network to pay more attention to high-precision positioning lights, reducing the impact of low-quality, low-precision positioning lights on the overall performance, and effectively improving the accuracy and robustness of the visible light positioning system.

[0043] 3. The present invention uses an LSTM network to model the temporal relationship between RSS values in adjacent positions of the same trajectory, so that the model can consider the RSS value over a period of time to estimate the current position, making up for the deficiency of graph neural networks that only consider spatial structure and information in the signal optimization process, and further improving the visible light positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a structural diagram of a visible light positioning system based on a photodiode in the prior art;

[0045] Figure 2 An overall structural diagram of a visible light positioning system based on a graph neural network provided by an embodiment of the present invention;

[0046] Figure 3 A flowchart of a visible light positioning method based on a graph neural network provided by an embodiment of the present invention;

[0047] Figure 4 A schematic diagram of converting received signal strength values into graph structure data provided by an embodiment of the present invention;

[0048] Figure 5 An LSTM unit diagram provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all structures.

[0050] Example

[0051] See also Figure 2 , Figure 2 This is a diagram of the overall structure of a visible light positioning system based on a graph neural network, provided in an embodiment of the present invention. The PD in the figure represents a photodiode. In this embodiment, the visible light signal is first optimized and noise is suppressed based on the graph neural network. Then, an attention mechanism is used to weight positioning lights based on distance and signal-to-noise ratio to guide the graph neural network's representation learning. Finally, spatiotemporal relationships are modeled based on an LSTM network to compensate for the graph neural network's limitation of only considering spatial structure.

[0052] See further Figure 3 , Figure 3 A flowchart of a visible light positioning method based on a graph neural network provided in an embodiment of the present invention, the method comprising:

[0053] S110. Based on the graph neural network guided by the attention mechanism, a visible light positioning model to be trained is constructed through a long short-term memory neural network.

[0054] Furthermore, the above S110 specifically includes:

[0055] S111. Based on the collected positioning training data, the spatial modeling capability of the graph neural network is utilized to optimize the positioning training data.

[0056] The positioning training data is the received signal strength value and corresponding position information of each light source in the visible light positioning system, and specifically includes the following steps:

[0057] Step 1: Collect the RSS values of each light source in the visible light positioning system and its corresponding position information. Sufficient training data is required. Figure 4 , training data can be collected dynamically by moving the robot (which can obtain real-time position) to reduce data collection costs. The robot carries a photodiode (PD), and the positioning light in this embodiment is an LED light.

[0058] Step 2: Construct the graph data structure.

[0059] Convert the collected RSS values with location tags into a graph data structure, such as Figure 4 Specifically, represents the collected training data, where Indicates the RSS value of the i-th positioning light at the j-th position, {px j ,py j} represents the position information of the jth point.

[0060] The graph structure is represented as G = (V, E, X), where V = {v i} contains n nodes, where v i represents the i-th node, E={e ik} consists of all edges, where e ik represents the edge between the i-th node and the k-th node, X={x i} is a set of feature vectors, where x i is the feature of the i-th node.

[0061] In the process of converting the visible light positioning system into a graph data structure, each f j will be converted into a graph G j Specifically, represents the i-th positioning light received by the photodiode at the j-th position, and n represents the total number of positioning lights sensed by the photodiode at the j-th position; if the photodiode senses the i-th positioning light and the k-th positioning light at the j-th position at the same time, then the edge e between the two nodes j ik =1 otherwise e j ik =0,x j i It represents the RSS value of the i-th positioning light sensed by the photodiode at the j-th position.

[0062] Step 3: Define the graph neural network structure. For example, the graph neural network structure in this embodiment includes two layers of GraphSage, where the first hidden layer has 16 neurons, and the second hidden layer also has 16 neurons. Both layers of the network sample a mean aggregator to aggregate neighbor information.

[0063] Step 4: Graph Neural Network Data Processing. The constructed graph structure data is fed into the graph neural network to optimize the RSS value. The graph neural network includes two steps: aggregation and node information update:

[0064] Step 4.1: The mean aggregator is calculated as follows,

[0065]

[0066] Where N(v) is all neighbor nodes of v, k = {1, 2} represents the k-th layer of the graph neural network. Initialize h0 to x i , which is the RSS value received by the photodiode. is the neighbor representation learned by the k-th layer neural network of the aggregator.

[0067] Step 4.2: Node information is updated as follows:

[0068]

[0069] CONCAT means directly concatenating two vectors, namely the representation vector of the v node's neighbors and the v node's own feature vector, W k is the weight matrix of the k-th layer graph network, and σ is the nonlinear mapping function. The complete graph neural network update method is implemented through the following algorithm steps:

[0070] 1. Input: Graph G = (V, E); input features W1 and W2 are network weight matrices; σ is a nonlinear function; mean aggregator MEAN is the average aggregation function; N(v) is the number of all neighbors of node v.

[0071] 2.

[0072] 3.for k={1,2}do

[0073] 4.for v∈Vdo

[0074] 5.

[0075] 6.

[0076] 7.end

[0077] 8.

[0078] 9.end

[0079] 10. Output: v∈V.

[0080] In this embodiment, step S111 uses graph neural networks for visible light positioning systems to perform signal optimization and noise suppression. Compared with the traditional method of only modeling noise in each lamp separately, graph neural networks can aggregate neighbor information and consider the topological structure of positioning lights in the visible light positioning system to learn more robust signal representations, thereby effectively suppressing noise in the signal and improving positioning accuracy.

[0081] S112. By assigning different weights to positioning lights through the attention mechanism, the graph neural network can focus more on high-precision light sources, thereby guiding the representation learning of the graph neural network.

[0082] In this embodiment, the above S112 specifically includes the following steps:

[0083] Step 1: Calculate the signal-to-noise ratio.

[0084] According to the Fast Fourier Transform (FFT) and the frequencies of different positioning lights, the corresponding RSS values are obtained, and the noise is obtained in other non-modulated frequency bands. The signal-to-noise ratio is calculated based on the obtained RSS values and noise.

[0085] Step 2: Design the attention network.

[0086] Optionally, the attention network in this embodiment consists of two fully connected layers (FC) and one SoftMax layer.

[0087] Step 3: Feed the SNR into the attention network to obtain the weights corresponding to different light sources. Specifically, combine the RSS values and the noise into a feature vector and feed it into the attention network.

[0088] Step 4: Weight the RSS values from different positioning lights. Multiply the weight output by the attention network with the corresponding RSS value to complete the weighting operation.

[0089] Specifically, attention is calculated as follows:

[0090]

[0091] Among them, h signal =[rss1,rss2,…,rss n ] and h noise=[noise1,noise2,…,noise M ] are the signal and noise vectors respectively.

[0092] The light source weight can be calculated as follows:

[0093]

[0094] a j is the weight of the j-th positioning light obtained by calculation;

[0095] The weighted value calculation formula of the positioning light is as follows:

[0096]

[0097] Among them, s j is the weighted RSS value, Indicates multiplication.

[0098] This embodiment uses the attention mechanism to guide the graph neural network to pay more attention to high-precision LED lights, thereby reducing the impact of positioning lights with low ranging accuracy on the overall performance of the visible light positioning system, thereby improving its robustness and positioning accuracy. Compared with traditional manual weighting methods or equal weighting methods, it is more accurate and effective.

[0099] S113. Model the temporal relationship between the received signal strength values in adjacent positions of the same trajectory based on the long short-term memory neural network (LSTM) to obtain a visible light positioning model to be trained.

[0100] Specifically, the above steps include:

[0101] Step 1: Collect weighted RSS values over a period of time.

[0102] Specifically, the weighted RSS is expressed as S={S t-l ,…,S t-1 ,S t}, where l is the sequence length, S t ={s1,s2,…,s n} is the weighted RSS representation received by the photodiode at the tth moment.

[0103] Step 2: Design the LSTM network structure.

[0104] The LSTM unit in this embodiment consists of a forget gate (f t ), input gate (i t ), output gate composition (o t ), whose structure is as follows Figure 5 In this embodiment, the LSTM hidden layer 1 contains 60 neurons.

[0105] Step 3: Send the weighted received signal strength value into the LSTM network structure.

[0106] Step 4: Save the output state of the last LSTM cell.

[0107] Among them, the LSTM calculation method is as follows:

[0108] f t =σ(w sf S t +w hf h t-1 +b f )

[0109] i t =σ(w si S t +w hi h t-1 +b i )

[0110] g t =tanh(w sg S t +w hg h t-1 +b g )

[0111] f t =f t c t-1 +i t g t

[0112] o t =σ(w so S t +w ho h t-1 +b o )

[0113] h t =o t tanh(c t )

[0114] Among them, w sf ,w hf ,w si ,w hi ,w sg ,w hg ,w so , and w ho are the weight matrices of the network, and their definitions are as follows. For example, w ho Represents the weight matrix between the hidden layer and the output gate. b f ,b i,b g , and b o is the corresponding bias.

[0115] This embodiment uses an LSTM network to model the temporal relationship between the RSS values of adjacent positions on the same trajectory, enabling the model to consider the RSS values over a period of time to estimate the current position. This makes up for the deficiency of graph neural networks that only consider spatial structure and information during signal optimization, and effectively improves the overall positioning performance.

[0116] S120: Train the visible light positioning model to be trained to obtain a trained visible light positioning model.

[0117] In this embodiment, after the visible light positioning model to be trained is constructed, the visible light positioning model to be trained is further trained to obtain a trained visible light positioning model. The model training process includes the following steps:

[0118] Step 1: Position prediction. The last LSTM hidden layer state is fed into the subsequent fully connected layer network to obtain the position information. The specific calculation method is as follows:

[0119]

[0120] Among them, w hl and b l is the weight matrix and bias of the fully connected layer. f() is a nonlinear function. is an estimate of the photodiode position.

[0121] Step 2: Define the loss function. The specific formula is as follows:

[0122]

[0123] Among them, (x t ,y t ) is the real position at the tth time slice, is the estimated position at the tth time slice, and m is the number of reference points.

[0124] Step 3: Calculate the residuals.

[0125] Step 4: Update the network parameters by back-propagating the error to obtain the trained visible light positioning model. In this embodiment, the Adam optimization algorithm is used to update the model network parameters.

[0126] S130: Input the collected real-time received signal strength values into the trained visible light positioning model to obtain real-time position information of the object to be positioned.

[0127] In this embodiment, after the visible light positioning model is trained, the real-time received signal strength values are collected for a period of time, the received signal strength values are fed into the trained visible light positioning model, and the position output by the model is returned to the user, so that the real-time position information of the object to be positioned can be obtained.

[0128] Furthermore, an embodiment of the present invention also provides a visible light positioning device based on a graph neural network, the device comprising:

[0129] A model building module, which is used to build a visible light positioning model to be trained using a long short-term memory neural network based on the graph neural network guided by the attention mechanism;

[0130] A model training module, used to train the visible light positioning model to be trained to obtain a trained visible light positioning model;

[0131] The real-time positioning module is used to input the collected real-time received signal strength values into the trained visible light positioning model to obtain the real-time position information of the object to be positioned.

[0132] Optionally, the model building module is specifically used to:

[0133] Based on the collected positioning training data, the positioning training data is optimized using the spatial modeling capability of the graph neural network;

[0134] The attention mechanism is used to assign different weights to the positioning lights, thereby guiding the representation learning of the graph neural network;

[0135] Based on the long short-term memory neural network (LSTM), the temporal relationship between the received signal strength values at adjacent positions on the same trajectory is modeled to obtain the visible light positioning model to be trained.

[0136] Among them, based on the collected positioning training data, a visible light positioning model based on graph neural network is constructed, including:

[0137] Dynamically collect the received signal strength value of the positioning light and its corresponding position information through the signal acquisition device;

[0138] Constructing the collected received signal strength values with location tags into a graph structure;

[0139] Define the graph neural network structure;

[0140] The graph data structure is input into the graph neural network structure for processing, and the received signal strength value is optimized.

[0141] Optionally, the graph neural network structure includes 2 layers of GraphSage, each hidden layer has 16 hidden neurons, and each layer of the network includes a mean aggregator.

[0142] Among them, the attention mechanism is used to assign different weights to the positioning lights, thereby guiding the representation learning of the graph neural network, including:

[0143] Determine the signal-to-noise ratio corresponding to each positioning light based on the received signal strength value and noise corresponding to each positioning light;

[0144] Build an attention network;

[0145] Inputting the signal-to-noise ratio into the attention network to obtain the weight corresponding to each positioning light;

[0146] A weighted operation is performed on the weight and the received signal strength value corresponding to each positioning light.

[0147] Optionally, the attention network consists of two fully connected layers and one SoftMax layer.

[0148] The long short-term memory (LSTM) neural network is used to model the temporal relationship between the received signal strength values at adjacent locations on the same trajectory, including:

[0149] Collect weighted received signal strength values over a period of time;

[0150] Build LSTM network structure;

[0151] The weighted received signal strength value is fed into the LSTM network structure, and the output state of the last LSTM cell is retained.

[0152] Optionally, the LSTM network structure consists of a forget gate, an input gate, and an output gate.

[0153] A visible light positioning device based on a graph neural network provided in an embodiment of the present invention can execute a visible light positioning method based on a graph neural network provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method, which will not be described in detail.

[0154] Note that the above are only preferred embodiments of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments, and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments and may include many other equivalent embodiments without departing from the concept of the present invention. The scope of the present invention is determined by the scope of the appended claims.

Claims

1. A visible light positioning method based on graph neural network, characterized in that: include: S110. Based on the graph neural network guided by the attention mechanism, a visible light positioning model to be trained is constructed through a long short-term memory neural network. S120: training the visible light positioning model to be trained to obtain a trained visible light positioning model; S130, inputting the collected real-time received signal strength value into the trained visible light positioning model to obtain real-time position information of the object to be positioned; The S110 includes: Based on the collected positioning training data, the positioning training data is optimized using the spatial modeling capability of the graph neural network; The attention mechanism is used to assign different weights to the positioning lights, thereby guiding the representation learning of the graph neural network; Based on the long short-term memory neural network (LSTM), the temporal relationship between the received signal strength values at adjacent positions on the same trajectory is modeled to obtain the visible light positioning model to be trained. The attention mechanism assigns different weights to the positioning lights, thereby guiding the representation learning of the graph neural network, including: Determine the signal-to-noise ratio corresponding to each positioning light based on the received signal strength value and noise corresponding to each positioning light; Build an attention network; Inputting the signal-to-noise ratio into the attention network to obtain the weight corresponding to each positioning light; A weighted operation is performed on the weight and the received signal strength value corresponding to each positioning light.

2. The method according to claim 1, characterized in that Based on the collected positioning training data, the spatial modeling capability of the graph neural network is utilized to optimize the positioning training data, including: Dynamically collect the received signal strength value of the positioning light and its corresponding position information through the signal acquisition device; Constructing the collected received signal strength values with location tags into a graph data structure; Define the graph neural network structure; The graph data structure is input into the graph neural network structure for processing, and the received signal strength value is optimized.

3. The method according to claim 2, characterized in that The graph neural network structure includes two layers of GraphSage, each hidden layer has 16 hidden neurons, and each layer of the network includes a mean aggregator.

4. The method according to claim 1, wherein The attention network consists of two fully connected layers and one SoftMax layer.

5. The method according to claim 1, wherein The long short-term memory (LSTM) neural network is used to model the temporal relationship between the received signal strength values at adjacent locations on the same trajectory, including: Collect weighted received signal strength values over a period of time; Build LSTM network structure; The weighted received signal strength value is fed into the LSTM network structure, and the output state of the last LSTM cell is retained.

6. The method according to claim 5, characterized in that The LSTM network structure consists of a forget gate, an input gate, and an output gate.

7. A visible light positioning device based on graph neural network, characterized in that: include: A model building module, which is used to build a visible light positioning model to be trained using a long short-term memory neural network based on the graph neural network guided by the attention mechanism; A model training module, used to train the visible light positioning model to be trained to obtain a trained visible light positioning model; A real-time positioning module is used to input the collected real-time received signal strength values into the trained visible light positioning model to obtain real-time position information of the object to be positioned; The model building module is specifically used to: Based on the collected positioning training data, the positioning training data is optimized using the spatial modeling capability of the graph neural network structure; The attention mechanism is used to assign different weights to the positioning lights, thereby guiding the representation learning of the graph neural network; Based on the long short-term memory (LSTM) neural network, the temporal relationship between the received signal strength values at adjacent positions on the same trajectory is modeled to obtain the visible light positioning model to be trained. The attention mechanism assigns different weights to the positioning lights, thereby guiding the representation learning of the graph neural network, including: Determine the signal-to-noise ratio corresponding to each positioning light based on the received signal strength value and noise corresponding to each positioning light; Build an attention network; Inputting the signal-to-noise ratio into the attention network to obtain the weight corresponding to each positioning light; A weighted operation is performed on the weight and the received signal strength value corresponding to each positioning light.

Images