Visible light communication and IMU fusion indoor positioning method based on adaptive dynamic neural network

Abstract

The invention discloses a visible light communication and IMU fusion indoor positioning method based on an adaptive dynamic neural network, and the method comprises the steps: building a space measurement coordinate system in an indoor measurement space where a plurality of anchor nodes are deployed, and collecting IMU data carried by a moving object under visible light communication and RSS data sent by all anchor nodes in the indoor space; the method comprises the following steps: establishing a plurality of data acquisition points in an indoor measurement space, independently storing all RSS data acquired by each acquisition point, storing IMU data acquired by each acquisition point, then preprocessing all data acquired by all acquisition points, and dividing the data into a training set and a test set; an ARDN network model is constructed, and the ARDN network model is trained and stored; a trained and stored ARDN network model is deployed into edge equipment through conversion of a deep learning framework, and positioning information of a moving object under different conditions is obtained in real time in a complex indoor environment.

Description

Technical Field

[0001] The invention relates to a visible light communication and IMU fusion indoor positioning method based on an adaptive dynamic neural network. Background Art

[0002] In recent years, indoor positioning technology has become a key factor in promoting the development of the industry and is widely used in robot navigation and AGV positioning. Since GPS signals are limited indoors, indoor positioning systems mainly rely on inertial sensors or small beacons such as Wi-Fi, Bluetooth, RFID, ultra-wideband and LED lights. Among them, Wi-Fi, Bluetooth and RFID are widely used due to their low cost and easy deployment, but their accuracy is limited; ultra-wideband has high accuracy but high cost. In contrast, visible light positioning (VLP) is receiving more and more attention due to its high accuracy, high update frequency and low cost. By modifying existing lighting facilities, the VLP system can provide real-time positioning in a variety of scenarios to assist robot navigation. At the same time, although technologies such as pedestrian detection (PDR) have been used for many years in indoor positioning calculations, the errors caused by the accumulated errors of the low-cost IMUs associated with them in positioning are still a major challenge.

[0003] In the field of indoor navigation, including AGV vehicle navigation and pedestrian navigation, most fusion methods are still based on traditional filtering algorithms, such as extended Kalman filter (EKF), particle filter (PF) and its variants. Although these methods are reasonable in theory, in practical applications, due to the difficulty of noise and error modeling in both line-of-sight and non-line-of-sight conditions in visible light communication signals in complex indoor environments, traditional fusion algorithms still face challenges in ensuring high-precision positioning, often resulting in low positioning accuracy and large errors. Therefore, how to effectively deal with the specific conditions of moving objects in complex indoor environments, confirm line-of-sight and non-line-of-sight conditions, and ensure high positioning accuracy, is still a problem that needs to be solved in current indoor positioning technology. Summary of the invention

[0004] The present invention is to solve the problems existing in the above-mentioned prior art and provide a visible light communication and IMU fusion indoor positioning method based on an adaptive dynamic neural network. On the basis of ensuring the positioning accuracy in a complex indoor environment, the present invention can dynamically determine in real time in a complex indoor environment, when a mobile object moves indoors, the location of the RSS data and IMU data collected under visible light communication, whether it is a full line-of-sight situation, a partial line-of-sight situation and a partial non-line-of-sight situation, or a full non-line-of-sight situation, and at the same time integrates multi-source data with reference value in different situations, dynamically distributes them to different sub-networks, and conducts training and testing, which significantly enhances the ability to improve positioning accuracy in a complex indoor environment.

[0005] The technical solutions adopted in the present invention are:

[0006] A visible light communication and IMU fusion indoor positioning method based on adaptive dynamic neural network includes the following steps:

[0007] Step 1: Establish a spatial measurement coordinate system in the indoor measurement space where several anchor nodes are deployed, collect the IMU data carried by the moving objects under visible light communication and the RSS data sent by all anchor nodes in the indoor space;

[0008] Step 2: Establish several data collection points in the indoor measurement space, save all RSS data collected by each collection point separately, and save the IMU data collected at each collection point, then pre-process all the data obtained by all collection points and divide them into training sets and test sets;

[0009] Step 3: Construct an ARDN network model, which includes an input layer, a dynamic discrimination layer, a dynamic network allocation layer, a fully connected layer and an output layer. The RSS data under visible light communication is input by the input layer, and the dynamic discrimination layer dynamically determines whether the data is line-of-sight or non-line-of-sight. The dynamic network allocation layer inputs the determined result into the sub-network corresponding to each result; the neurons corresponding to the output of the sub-network output the positioning information of the moving object through the fully connected layer and the output layer;

[0010] Step 4: Train and save the ARDN network model;

[0011] Step 5: The trained and saved ARDN network model is deployed to the edge device through conversion of the deep learning framework to obtain the positioning information of moving objects in different situations in real time in complex indoor environments.

[0012] Furthermore, in step 1, a spatial measurement coordinate system XYZ is established in an indoor measurement space with multiple lighting devices, each lighting device is used as an anchor node, and the installation position (x i ′,y i ′,z i ′);

[0013] The RSS data of each anchor node is received through the photodiode on the mobile object. The RSS data includes the size and ID information of the anchor node RSS data value.

[0014] Furthermore, several data collection points are established in the indoor measurement space, and the RSS data of all anchor nodes collected by each collection point are saved separately to obtain the RSS data at each collection point; the 9-axis IMU data collected at each collection point are saved separately to obtain 9 groups of data at each collection point;

[0015] All data acquired from all acquisition points are standardized and normalized, and combined with the corresponding timestamps and acquisition point locations (x i ,y i ,z i ) as the data set, 80% of the total batches are taken as the training set and the rest are the test set.

[0016] Furthermore, the dynamic judgment layer is used to judge the stability of the RSS signal of each anchor node under visible light communication, and establish a judgment threshold τ of the dynamic judgment layer. By comparing the collected RSS value with the threshold τ, when it is lower than the threshold τ, the optical signal of the corresponding anchor node is blocked by an obstacle, and the anchor node is judged to be non-line-of-sight; when it is higher than or equal to the threshold τ, the optical signal of the corresponding anchor node is not blocked by an obstacle, and the anchor node is judged to be line-of-sight.

[0017] Furthermore, the dynamic network allocation layer performs dynamic allocation according to the result of the dynamic judgment layer, and the allocation is as follows:

[0018] When the number of anchor nodes is even, the dynamic network allocation layer allocates data to the corresponding subnetwork in the following way:

[0019] When in full line of sight, all RSS data and IMU data are input into the first sub-network model for positioning prediction;

[0020] When both line-of-sight and non-line-of-sight situations occur, when the line-of-sight RSS number minus the non-line-of-sight RSS number is greater than or equal to 2, the RSS data and IMU data under the line-of-sight situation are input into the second sub-network model for positioning prediction;

[0021] When the number of line-of-sight RSS is equal to the number of non-line-of-sight RSS, the RSS data and IMU data under the two line-of-sight conditions are input into the third sub-network model for positioning prediction;

[0022] When the non-line-of-sight RSS number minus the line-of-sight RSS number is greater than or equal to 2, the line-of-sight RSS data and IMU data are input into the fourth sub-network model for positioning prediction;

[0023] In the case of non-line-of-sight, the IMU data is input into the fifth sub-network model for positioning prediction;

[0024] When the number of anchor nodes is odd, the dynamic network allocation layer allocates data to the corresponding subnetwork in the following way:

[0025] When in full line of sight, all RSS data and IMU data are input into the first sub-network model for positioning prediction;

[0026] When both line-of-sight and non-line-of-sight situations occur, when the line-of-sight RSS number minus the non-line-of-sight RSS number is greater than or equal to 2, the RSS data and IMU data under the line-of-sight situation are input into the second sub-network model for positioning prediction;

[0027] When the absolute value of the difference between the line-of-sight RSS number and the non-line-of-sight RSS number is 1, the line-of-sight RSS data and IMU data are input into the third sub-network model for positioning prediction;

[0028] When the non-line-of-sight RSS number minus the line-of-sight RSS number is greater than or equal to 2, the line-of-sight RSS data and IMU data are input into the fourth sub-network model for positioning prediction;

[0029] In a completely non-line-of-sight situation, the IMU data is input into the fifth sub-network model for positioning prediction.

[0030] Furthermore, the first sub-network model is a single LSTM network;

[0031] The second sub-network model is a complex LSTM network;

[0032] The third sub-network model is the CNN-LSTM network;

[0033] The fourth sub-network model is the CNN-LSTM-Attention network;

[0034] The fifth sub-network model is the CNN-BiLSTM-Attention network.

[0035] Furthermore, the single LSTM network includes 2 hidden layers and 1 fully connected layer, and the hidden layer output is 16 LSTM neurons;

[0036] The complex LSTM network consists of 2 hidden layers and 1 fully connected layer, and the hidden layer output is 32 LSTM neurons;

[0037] The CNN-LSTM network consists of 1 convolutional layer, 1 pooling layer, 2 LSTM layers, and 1 fully connected layer;

[0038] The CNN-LSTM-Attention network consists of 1 convolutional layer, 1 pooling layer, 2 LSTM layers, 1 attention layer, and 1 fully connected layer;

[0039] The CNN-BiLSTM-Attention network consists of 1 convolutional layer, 1 pooling layer, 2 forward and reverse LSTM layers, 1 attention layer and 1 fully connected layer.

[0040] The present invention has the following beneficial effects:

[0041] When collecting data and IMU data under visible light communication for mobile objects in complex indoor environments, line-of-sight and non-line-of-sight situations may occur. How to identify the specific situations of line-of-sight and non-line-of-sight is particularly important. The method of the present invention utilizes the establishment of an adaptive RSS-IMU dynamic neural network, which can well distinguish the specific situation of the position of the mobile object in a complex indoor environment, and dynamically allocate it to different sub-networks according to different situations. It combines multiple data sources, effectively processes and fuses these data through a neural network model, and performs positioning prediction, which significantly enhances the ability to improve positioning accuracy in complex indoor environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of the present invention.

[0043] Figure 2 A schematic diagram of an indoor scene setting according to an embodiment of the present invention;

[0044] Figure 3 It is a schematic diagram of the deep learning network model used in the implementation of the present invention. DETAILED DESCRIPTION

[0045] like Figure 1 and Figure 2 As shown, this embodiment takes four lighting LEDs as an example (i.e. Figure 2 The method of the present invention is described as follows with reference to LED1, LED2, LED3 and LED4 in three-dimensional space.

[0046] The steps of the present invention are as follows:

[0047] Step 1: Set up the reference coordinate system. Establish a spatial measurement coordinate system XYZ in the indoor measurement space. Set up four lighting devices (i.e. LEDs) as anchor nodes in the indoor environment and record the installation positions (x i ′,y i ′,z i ′), the illumination coverage area of ​​the lighting device can cover the moving range of the mobile object, collect the IMU positioning measurement data carried by the mobile object, and receive the RSS data of the four anchor nodes under the visible light communication through the photodiode (PD) on the mobile object. The RSS data includes the size and ID information of the RSS data value of the anchor node.

[0048] Step 2: Establish several data collection points in the indoor measurement space, save the RSS data of the four anchor nodes collected by each collection point in four arrays, and obtain four groups of data. Save the 9-axis IMU data collected at each collection point separately, and obtain nine groups of data. Standardize and normalize all the acquired data, and combine the corresponding timestamp and collection point position (x i ,yi ,z i ) as the data set, taking 80% of the total batch as the training set and the rest as the test set;

[0049] Step 3: Establish an Adaptive RSS-IMU Dynamic Neural Network (ARDN) model. The network model contains an input layer, a dynamic discrimination layer, a dynamic network allocation layer, a fully connected layer, and an output layer. The network model can adaptively select the appropriate sub-network for processing according to the different situations of the input data, use the dynamic discrimination layer to judge the fluctuation of the input data RSS signal, dynamically allocate sub-networks to perform processing tasks, and then output the positioning information in complex indoor environments through the fully connected layer and the output layer.

[0050] Step 4: Train and save the ARDN network model.

[0051] The activation function used in the hidden layer of each sub-network in the ARDN network model is Relu. The optimizer of the network model uses the Adam method, the loss function uses MSE, and the initial learning rate is set to 0.001. Input the training data to train the dynamic network model, and save it to the specified path after the training is completed.

[0052] Step 5: Model deployment to achieve real-time positioning.

[0053] When deploying edge devices, it is necessary to convert the deep learning framework. Obtain the original parameters of the ARDN network model, compare them with the model parameters of the existing MindSpore framework, load the model of the original TensorFlow framework to the position corresponding to the MindSpore parameter name through the key name mapping table, and then use MindSpore to load the corresponding parameters, and deploy them to the edge device Hongmeng development board through the tool DevEco Studio. On the Hongmeng development board, the real-time visible light positioning data and IMU data received are used as input to obtain positioning information in a complex indoor environment.

[0054] Step 1 is as follows:

[0055] Four anchor nodes are deployed in the indoor positioning space, and the LED light signal is modulated by on-off keying (OOK). The control information sent by the anchor node in the visible light communication system usually contains its coordinate information and ID information. Through the indoor photoelectric receiver (Photo Diode, PD), at a certain collection point, the intensity values ​​{RSSI1, RSSI2, RSSI3, RSSI4} of different anchor nodes under visible light communication are received. Multiple sets of data are collected at the same collection point, and the collection point coordinates (x i ,y i,z i ), initially calibrate the IMU data carried by the moving object, and collect the three-axis acceleration data acc_x, acc_y, acc_z, the three-axis angular velocity gyro_x, gyro_y, gyro_z and the three-axis magnetic force mag_x, mag_y, mag_z of the moving object during movement.

[0056] The details of step 2 are as follows:

[0057] Save the timestamp of each acquisition point, as well as the 4 sets of RSS data, 9 sets of IMU data and the specific coordinates of each acquisition point. Collect enough data. Assuming that 1000 acquisition points are deployed in the indoor environment, the specific data of 1000 acquisition points can be collected. These 1000 acquisition points cover the indoor places without shadows (i.e., full line of sight), the junction of shadows and non-shadows (i.e., the situation containing line of sight and non-line of sight), and the full shadows (i.e., the situation of full non-line of sight). Take 80% of the total batch to form the training set, and the remaining 20% ​​to form the test set. The training set contains train_x and train_y, and the test set includes test_x and test_y.

[0058] The Adaptive RSS-IMU Dynamic Neural Network (ARDN) model in the present invention is as follows: Figure 3 As shown, the details are as follows:

[0059] The Adaptive RSS-IMU Dynamic Neural Network (ARDN) model contains an input layer, a dynamic discrimination layer, a dynamic network allocation layer, a fully connected layer, and an output layer.

[0060] The input features of the input layer include the timestamp Timaestamp, the intensity values ​​of visible light communication under the four anchor nodes {RSSI1, RSSI2, RSSI3, RSSI4}, the nine-axis sensor data of the IMU: including three-axis acceleration data acc_x, acc_y, acc_z, three-axis angular velocity gyro_x, gyro_y, gyro_z and three-axis magnetic force mag_x, mag_y, mag_z, and the real coordinates of the collection point, a total of 17 input features.

[0061] The dynamic discrimination layer can judge the stability of the RSS signal under visible light communication. By setting a suitable threshold τ, the RSS value received under visible light communication at the four anchor nodes is compared with the threshold τ. It is judged that in a complex indoor environment, the RSS under visible light communication is received by the indoor photoelectric receiver (Photo Diode, PD). It is judged that the RSS value received at the four anchor nodes is lower than the threshold τ, that is, the moving object is at the collection point and the relevant anchor nodes are blocked by obstacles (that is, partial non-line-of-sight occurs). When there are 4 anchor nodes in the indoor environment, there will be 5 situations, namely full line-of-sight, 1 non-line-of-sight and 3 line-of-sight, 2 non-line-of-sight and 2 line-of-sight, 3 non-line-of-sight and 1 line-of-sight, and 4 full non-line-of-sight.

[0062] The dynamic network allocation layer is based on the results of the dynamic discrimination layer. It dynamically allocates the data to five different sub-networks for processing according to the specific fluctuations of the RSS values ​​received under the four anchor nodes under visible light communication. It also uses the visible light RSS data and IMU data in a stable state, combined with features such as timestamps and actual coordinates of the collection points, and inputs them into the corresponding sub-network models.

[0063] When the number of anchor nodes is an even number, take 4 anchor nodes in this example.

[0064] When the moving object is in full line of sight, all RSS data and IMU data are input into the first sub-network model for positioning prediction.

[0065] When both line-of-sight and non-line-of-sight occur at the same time, and the line-of-sight RSS number minus the non-line-of-sight RSS number is greater than or equal to 2, that is, when the moving object is in 1 non-line-of-sight and 3 line-of-sight situations in this example, the RSS data and IMU data in the 3 line-of-sight situations are input into the second sub-network model for positioning prediction.

[0066] When the number of line-of-sight RSS is equal to the number of non-line-of-sight RSS, that is, when the moving object is in 2 non-line-of-sight and 2 line-of-sight situations in this example, the RSS data and IMU data in the 2 line-of-sight situations are input into the third sub-network model for positioning prediction.

[0067] When the non-line-of-sight RSS number minus the line-of-sight RSS number is greater than or equal to 2, that is, when the moving object is in 3 non-line-of-sight and 1 line-of-sight situations in this example, the RSS data and IMU data in the 1 line-of-sight situation are input into the fourth sub-network model for positioning prediction.

[0068] When the moving object is in a completely non-line-of-sight situation, the IMU data is input into the fifth sub-network model for positioning prediction.

[0069] When the number of anchor nodes is an odd number, take 7 anchor nodes as an example.

[0070] When the moving object is in full line of sight, all RSS data and IMU data are input into the first sub-network model for positioning prediction.

[0071] When both line-of-sight and non-line-of-sight occur at the same time, and the line-of-sight RSS number minus the non-line-of-sight RSS number is greater than or equal to 2, that is, there are 1 non-line-of-sight and 6 line-of-sight situations and 2 non-line-of-sight and 5 line-of-sight situations, the RSS data and IMU data in the line-of-sight situation are input into the second sub-network model for positioning prediction.

[0072] When the absolute value of the difference between the number of line-of-sight RSS and the number of non-line-of-sight RSS is 1, that is, there are 3 non-line-of-sight and 4 line-of-sight cases and 4 non-line-of-sight and 3 line-of-sight cases, the RSS data and IMU data in the line-of-sight case are input into the third sub-network model for positioning prediction.

[0073] When the number of non-line-of-sight RSS minus the number of line-of-sight RSS is greater than or equal to 2, that is, there are 5 non-line-of-sight and 2 line-of-sight cases and 6 non-line-of-sight and 1 line-of-sight case. The RSS data and IMU data in the line-of-sight case are input into the fourth sub-network model for positioning prediction.

[0074] When the moving object is in a completely non-line-of-sight situation, the IMU data is input into the fifth sub-network model for positioning prediction.

[0075] When the number of anchor nodes is even or odd, the data collected at each collection point is input into different sub-networks when different situations occur. Different sub-networks output corresponding features and then pass through the fully connected layer and the output layer to output the corresponding positioning information.

[0076] The first sub-network model is a single LSTM network, which includes 2 hidden layers and 1 fully connected layer, and the hidden layer output is 16 LSTM neurons.

[0077] The second sub-network model is a complex LSTM network, which contains 2 hidden layers and 1 fully connected layer, and the hidden layer output is 32 LSTM neurons.

[0078] The third sub-network model is a CNN-LSTM network, which includes 1 convolutional layer, 1 pooling layer, 2 LSTM layers and 1 fully connected layer.

[0079] The fourth sub-network model is the CNN-LSTM-Attention network, which includes 1 convolutional layer, 1 pooling layer, 2 LSTM layers, 1 attention layer and 1 fully connected layer.

[0080] The fifth sub-network model is the CNN-BiLSTM-Attention network, which includes 1 convolutional layer, 1 pooling layer, 2 forward and reverse LSTM layers, 1 attention layer and 1 fully connected layer.

[0081] After passing through the dynamic discrimination layer and the dynamic network allocation layer, the output features of the fully connected layer in the corresponding subnetwork are input into the fully connected layer in the Adaptive RSS-IMU Dynamic Neural Network (ARDN) model, and then the specific positioning information in the complex indoor environment is output in multiple dimensions through the softmax function.

[0082] In this example, the implementation process of step 4 is as follows:

[0083] The trained ARDN network model is saved. The activation function used in the hidden layer of each sub-network in the model is Relu, the loss function is MSE, and the number of iterations Epoch is set to 30.

[0084] The activation function Relu formula of the hidden layer:

[0085] Relu(x)=max(0,x)

[0086] Loss function MSE formula:

[0087]

[0088] where y i is the actual value of the sampling point, y i ′ is the predicted value of the ARDN network model, n is the total number of batches of data, and a total of 30 training times are conducted. The training adopts the Tensorflow deep learning framework. After the training is completed, the model is saved to the specified path.

[0089] The implementation process of step 5 is as follows:

[0090] From the trained ARDN network model file, obtain the parameters of the ARDN network model under the TensorFlow deep learning framework, and convert the trained TensorFlow model into a checkpoint that can be used by MindSpore.

[0091] Print the parameter names and shapes of all parameters in the TensorFlow parameter file, and print the parameter names and shapes of all parameters in the MindSpore Cell that needs to load the parameters.

[0092] Compare parameter name and shape and construct parameter mapping relationship.

[0093] Map TensorFlow parameters to MindSpore parameters according to parameter mapping, form a Parameter List and save it as a checkpoint.

[0094] MindSpore loads the converted parameters, fixes the input, and compares the results of MindSpore and TensorFlow to see whether the difference between the results is large. If the difference is large, further model optimization is required.

[0095] Deploy to the edge device Hongmeng development board through the tool DevEco Studio.

[0096] On the Hongmeng development board, the RSS data and IMU data received in real time under visible light communication are used as input to obtain real-time positioning information in complex indoor environments.

[0097] The above embodiments are preferred implementations of the present invention, but the present invention is not limited to the above specific implementations. Within the technical concept of the present invention, the present invention can be subjected to a variety of simple modifications, and these simple modifications all belong to the protection scope of the present invention. In the above embodiments, four anchor nodes are arranged in the indoor environment, which can meet the positioning accuracy requirements of the present invention. However, the technical solution of the present invention is not limited to this. Without departing from the spirit of the present invention, five or more anchor nodes can also be arranged in the indoor environment. After increasing the number of anchor nodes, the system described in the present invention can still be used normally, and only the dynamic discrimination layer and the dynamic network allocation layer need to be adjusted accordingly.

[0098] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be regarded as within the protection scope of the present invention.

Claims

1. A visible light communication and IMU fusion indoor positioning method based on adaptive dynamic neural network, characterized by: The following steps are involved: Step 1: Establish a spatial measurement coordinate system in the indoor measurement space where several anchor nodes are deployed, collect the IMU data carried by the moving objects under visible light communication and the RSS data sent by all anchor nodes in the indoor space; Step 2: Establish several data collection points in the indoor measurement space, save all RSS data collected by each collection point separately, and save the IMU data collected at each collection point, then pre-process all the data obtained by all collection points and divide them into training sets and test sets; Step 3: Construct an ARDN network model, which includes an input layer, a dynamic discrimination layer, a dynamic network allocation layer, a fully connected layer and an output layer. The RSS data under visible light communication is input by the input layer, and the dynamic discrimination layer dynamically determines whether the data is line-of-sight or non-line-of-sight. The dynamic network allocation layer inputs the determined result into the sub-network corresponding to each result; the neurons corresponding to the output of the sub-network output the positioning information of the moving object through the fully connected layer and the output layer; Step 4: Train and save the ARDN network model; Step 5: The trained and saved ARDN network model is deployed to the edge device through conversion of the deep learning framework to obtain the positioning information of moving objects in different situations in real time in complex indoor environments.

2. The indoor positioning method based on visible light communication and IMU fusion based on adaptive dynamic neural network as claimed in claim 1, characterized in that: In step 1, a spatial measurement coordinate system XYZ is established in an indoor measurement space with multiple lighting devices. Each lighting device is used as an anchor node to record the installation position (x i ′,y i ′,z i ′); The RSS data of each anchor node is received through the photodiode on the mobile object. The RSS data includes the size and ID information of the anchor node RSS data value.

3. The indoor positioning method based on visible light communication and IMU fusion based on adaptive dynamic neural network as claimed in claim 2, characterized in that: Establish several data collection points in the indoor measurement space, save the RSS data of all anchor nodes collected by each collection point separately, and obtain the RSS data at each collection point; The 9-axis IMU data collected at each collection point are saved separately to obtain 9 sets of data at each collection point; All data acquired from all acquisition points are standardized and normalized, and combined with the corresponding timestamps and acquisition point locations (x i ,y i ,z i ) as the data set, 80% of the total batches are taken as the training set and the rest are the test set.

4. The indoor positioning method based on visible light communication and IMU fusion based on adaptive dynamic neural network as claimed in claim 2, characterized in that: The dynamic discrimination layer is used to judge the stability of the RSS signal of each anchor node under visible light communication, and to set a discrimination threshold τ of the dynamic discrimination layer. By comparing the collected RSS value with the threshold τ, when the value is lower than the threshold τ, the optical signal of the corresponding anchor node is blocked by an obstacle, and the anchor node is judged to be non-line-of-sight; when the value is higher than or equal to the threshold τ, the optical signal of the corresponding anchor node is not blocked by an obstacle, and the anchor node is judged to be line-of-sight.

5. The indoor positioning method based on visible light communication and IMU fusion based on adaptive dynamic neural network as claimed in claim 4, characterized in that: The dynamic network allocation layer performs dynamic allocation based on the results of the dynamic judgment layer. The allocation is as follows: When the number of anchor nodes is even, the dynamic network allocation layer allocates data to the corresponding subnetwork in the following way: When in full line of sight, all RSS data and IMU data are input into the first sub-network model for positioning prediction; When both line-of-sight and non-line-of-sight situations occur, when the line-of-sight RSS number minus the non-line-of-sight RSS number is greater than or equal to 2, the line-of-sight RSS data and IMU data are input into the second sub-network model for positioning prediction; When the number of line-of-sight RSS is equal to the number of non-line-of-sight RSS, the RSS data and IMU data under the two line-of-sight conditions are input into the third sub-network model for positioning prediction; When the non-line-of-sight RSS number minus the line-of-sight RSS number is greater than or equal to 2, the line-of-sight RSS data and IMU data are input into the fourth sub-network model for positioning prediction; In the case of non-line-of-sight, the IMU data is input into the fifth sub-network model for positioning prediction; When the number of anchor nodes is odd, the dynamic network allocation layer allocates data to the corresponding subnetwork in the following way: When in full line of sight, all RSS data and IMU data are input into the first sub-network model for positioning prediction; When both line-of-sight and non-line-of-sight situations occur, when the line-of-sight RSS number minus the non-line-of-sight RSS number is greater than or equal to 2, the line-of-sight RSS data and IMU data are input into the second sub-network model for positioning prediction; When the absolute value of the difference between the line-of-sight RSS number and the non-line-of-sight RSS number is 1, the line-of-sight RSS data and IMU data are input into the third sub-network model for positioning prediction; When the non-line-of-sight RSS number minus the line-of-sight RSS number is greater than or equal to 2, the line-of-sight RSS data and IMU data are input into the fourth sub-network model for positioning prediction; In a completely non-line-of-sight situation, the IMU data is input into the fifth sub-network model for positioning prediction.

6. The indoor positioning method based on visible light communication and IMU fusion based on adaptive dynamic neural network as claimed in claim 5, characterized in that: Said The first sub-network model is a single LSTM network; The second sub-network model is a complex LSTM network; The third sub-network model is the CNN-LSTM network; The fourth sub-network model is the CNN-LSTM-Attention network; The fifth sub-network model is the CNN-BiLSTM-Attention network.

7. The indoor positioning method based on visible light communication and IMU fusion based on adaptive dynamic neural network as claimed in claim 6, characterized in that: The single LSTM network consists of 2 hidden layers and 1 fully connected layer, and the hidden layer output is 16 LSTM neurons; The complex LSTM network consists of 2 hidden layers and 1 fully connected layer, and the hidden layer output is 32 LSTM neurons; The CNN-LSTM network consists of 1 convolutional layer, 1 pooling layer, 2 LSTM layers, and 1 fully connected layer; The CNN-LSTM-Attention network consists of 1 convolutional layer, 1 pooling layer, 2 LSTM layers, 1 attention layer, and 1 fully connected layer; The CNN-BiLSTM-Attention network consists of 1 convolutional layer, 1 pooling layer, 2 forward and reverse LSTM layers, 1 attention layer and 1 fully connected layer.

Images