A BQN indoor positioning method based on multi-data fusion

By integrating Wi-Fi and pedestrian dead information in the width network and optimizing parameters with ridge regression, the problems of low indoor positioning efficiency and model training efficiency in the existing technology are solved, and efficient indoor positioning and model training are achieved.

CN118836870BActive Publication Date: 2024-11-26CENT SOUTH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411310941.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-20
Publication Date
2024-11-26
Estimated Expiration
2044-09-20

AI Technical Summary

Technical Problem

The existing deep reinforcement learning indoor positioning methods have shortcomings in multi-data fusion and model training efficiency, and the simulation effect of deep networks is not good, and the network needs to be rebuilt to improve performance.

Method used

The width network is used to combine the Wi-Fi receiving signal strength of pedestrian trajectory and the heading angle of pedestrian dead reckoning. The action is selected through greedy strategies, and the network parameters are calculated using the ridge regression method. The model fitting ability is improved through incremental learning and the network model is optimized. To improve positioning efficiency.

Benefits of technology

While maintaining good positioning performance, it significantly reduces model training time and improves training efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118836870B_ABST
    Figure CN118836870B_ABST
Patent Text Reader

Abstract

The embodiment of the present invention provides a multi-data fusion BQN indoor positioning method, which belongs to the field of measurement technology, and specifically includes: inputting the state of the intelligent body into the width network; the width network extracts features of the state of the intelligent body, outputs an action value vector, and selects an action according to the #imgabs0#-greedy strategy; calculating rewards and storing experience samples in the experience pool, randomly extracting experience samples from the experience pool, and calculating the action value vector according to the corresponding action output of the width network; using the ridge regression method to calculate the parameters of the width network, if the fitting accuracy is not met, the fitting ability of the model is improved by incremental learning; the real-time collected Wi-Fi received signal strength indication and the heading angle estimated by pedestrian dead reckoning are input into the trained parameter optimization network model, and the action with the highest value is selected from the action vector output by the width network to estimate the pedestrian position. Through the scheme of the present invention, the positioning efficiency and model training efficiency are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of measurement technology, and in particular to a BQN indoor positioning method with multi-data fusion. Background Art

[0002] With the continuous development of smart city services, the demand for precise indoor positioning is growing, especially in the context that the Global Positioning System (GPS) often fails to provide reliable positioning in indoor environments. The popularity of smartphones has provided new opportunities for indoor positioning technology, especially the method of combining Wi-Fi fingerprint library positioning has shown its importance in various application scenarios.

[0003] At present, most deep reinforcement learning indoor positioning methods are designed based on deep Q-network deep reinforcement learning algorithms, using single Wi-Fi or pedestrian dead reckoning (PDR) information for positioning. Although these methods have achieved good positioning results, they ignore the effect of multi-data fusion. In addition, these methods use deep networks to fit the Q function. Since deep networks contain a large number of connection parameters, it takes a lot of time to train positioning models using a large amount of pedestrian data. In addition, if the simulation effect of the deep network is not good, researchers need to rebuild the network to improve performance.

[0004] It can be seen that there is an urgent need for a multi-data fusion BQN indoor positioning method that can both effectively locate and improve the efficiency of training models. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides a multi-data fusion BQN indoor positioning method, which at least partially solves the problems of poor positioning efficiency and model training efficiency in the prior art.

[0006] The embodiment of the present invention provides a multi-data fusion BQN indoor positioning method, including:

[0007] Step 1: Use the Wi-Fi received signal strength indication of the pedestrian trajectory and the heading angle estimated by pedestrian dead reckoning as the first Step agent state Input to the width network;

[0008] Step 2: The wide network extracts features from the agent’s state and outputs an action value vector ,according to - Greedy strategy to select actions ;

[0009] Step 3: Calculate rewards And the experience sample Store in the experience pool and randomly extract experience samples from the experience pool , calculate the action value vector based on the corresponding action output of the width network;

[0010] Step 4: Calculate the parameters of the width network using the ridge regression method. If the fitting accuracy is not met, improve the fitting ability of the model through incremental learning to obtain a trained parameter optimization network model.

[0011] In step 5, the real-time collected Wi-Fi received signal strength indication and the heading angle estimated by pedestrian dead reckoning are input into the trained parameter optimization network model, and the action with the highest value is selected from the action vector output by the width network to estimate the pedestrian position.

[0012] According to a specific implementation of the embodiment of the present invention, the -The expression of the greedy strategy is

[0013] ;

[0014] in, Represents the vector used to find the action value The function of the action when the action Q value reaches the maximum value.

[0015] According to a specific implementation of the embodiment of the present invention, step 3 specifically includes:

[0016] Step 3.1, calculate the reward according to the reward function, where the expression of the reward function is

[0017]

[0018] Among them, the agent is in position Execute an action , move to position , Indicates that the agent is at position The target position, dis function is used to calculate the distance between two positions;

[0019] Step 3.2, from the experience pool Randomly select a certain amount of samples , based on which the target value is calculated , where the target value The expression is

[0020] ;

[0021] in is based on The maximum action output by the target Q network, is the target Q function;

[0022] Step 3.3, calculate The action value vector , where the action value vector The expression is

[0023]

[0024] in, is a vector representing the Q value of each action, G represents all the actions of the agent, It is the Q function, which is used to calculate the Q value of the action;

[0025] Step 3.4, Corresponding to the action The Q value is to update.

[0026] According to a specific implementation of the embodiment of the present invention, step 4 specifically includes:

[0027] Step 4.1, change the status As input, it is fed into the left feature node of the first layer, where the expression of the left feature node is

[0028] ;

[0029] Among them, each feature group has a weight matrix and the bias matrix , whose coefficients are randomly generated, is a nonlinear function, the output of each feature node Cascaded , expressed as ;

[0030] Step 4.2, Input into the enhancement group to get the enhanced node output ;

[0031] ;

[0032] in, is the weight matrix, is the bias matrix, and They are all randomly generated. is a nonlinear function;

[0033] Step 4.3: Output the feature group And the enhanced group output is spliced ​​to get , expressed as ;

[0034] Step 4.4, using the ridge regression method, calculate The pseudo-inverse of ,

[0035] ;

[0036] in, is the ridge coefficient, Represents the identity matrix.

[0037] Step 4.5: If the fitting error of width learning exceeds the set threshold, an enhancement node is added, and the additional enhancement node generates a new , the expression is , and are the new weight matrix and bias matrix respectively, and then let , calculated according to the pseudo-inverse theory of block matrices ,

[0038]

[0039] in, Expressed as ,

[0040] ;

[0041] Step 4.6, for The increment of , calculates the new parameters of the width network ,

[0042] ;

[0043] Step 4.7, each time step 1 to step 4.6 is repeated for a preset number of times, the Q network parameters are synchronized with the target Q network parameters, and training is performed until convergence to obtain a trained parameter optimization network model.

[0044] According to a specific implementation of the embodiment of the present invention, step 5 specifically includes:

[0045] The real-time Wi-Fi received signal strength indication and the heading angle estimated by pedestrian dead reckoning are input into the model. The agent selects an action in the model output according to the highest Q value and performs this action in combination with the moving distance estimated by pedestrian dead reckoning, thereby transferring to the next state. This process is repeated to obtain the agent's trajectory and estimate the pedestrian's path.

[0046] The BQN indoor positioning solution of multiple data fusion in the embodiment of the present invention includes: step 1, using the Wi-Fi received signal strength indication of the pedestrian trajectory and the heading angle estimated by the pedestrian dead reckoning as the first Step agent state Input to the wide network; Step 2, the wide network extracts features of the agent's state and outputs the action value vector ,according to - Greedy strategy to select actions ; Step 3, calculate the reward And the experience sample Store in the experience pool and randomly extract experience samples from the experience pool , calculate the action value vector according to the corresponding action output of the width network; step 4, use the ridge regression method to calculate the parameters of the width network. If the fitting accuracy is not met, improve the fitting ability of the model through incremental learning to obtain a trained parameter optimization network model; step 5, input the real-time collected Wi-Fi received signal strength indication and the heading angle estimated by pedestrian dead reckoning into the trained parameter optimization network model, and select the action with the highest value from the action vector output by the width network to estimate the pedestrian position.

[0047] The beneficial effect of the embodiment of the present invention is that: through the scheme of the present invention, the Wi-Fi received signal strength indication of the pedestrian trajectory and the heading angle estimated by the pedestrian dead reckoning are used as the state of the intelligent body. Input to the wide network; the wide network extracts features of the agent's state and outputs an action value vector , the agent according to - Greedy strategy to select actions ; Calculate rewards and put experience samples Store in the experience pool. Randomly extract experience samples from the experience pool , calculate the action value vector according to the corresponding action output of the target width network; use the ridge regression method to calculate the parameters of the width network. If the fitting accuracy is not met, improve the fitting ability of the model through incremental learning to obtain the trained parameter optimization network model; input the real-time collected Wi-Fi received signal strength indication and the heading angle estimated by pedestrian dead reckoning into the trained parameter optimization network model, select the action with the highest value from the action vector output by the width network to estimate the pedestrian position, and convert the parameter optimization method based on the deep network into the parameter optimization method based on the width network. Compared with the existing technology, it maintains good positioning performance while reducing the training time of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0049] Figure 1 A schematic diagram of a flow chart of a BQN indoor positioning method with multi-data fusion provided by an embodiment of the present invention;

[0050] Figure 2 A structural diagram of a parameter optimization network model provided by an embodiment of the present invention;

[0051] Figure 3 A model training architecture diagram provided for an embodiment of the present invention;

[0052] Figure 4 A comparison diagram of the positioning effects of the method of the present invention and the prior art provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0054] The following describes the embodiments of the present invention through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0055] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present invention, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this device and / or practice this method.

[0056] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention. The drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0057] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.

[0058] The common features of Wi-Fi positioning and PDR positioning are low cost and no need to deploy special equipment support, which is very suitable for indoor positioning of smartphones. Among them, Wi-Fi signals are easily interfered by the indoor environment, resulting in low accuracy of Wi-Fi-based positioning methods. While the PDR positioning method can provide high-precision positioning in the short term, it will be affected by the noise of the built-in sensors of smartphones and there will be error accumulation, while Wi-Fi positioning is not affected by error accumulation. The positioning method based on the fusion of Wi-Fi and PDR combines the characteristics of Wi-Fi positioning and PDR positioning. Wi-Fi and PDR complement each other. Wi-Fi overcomes the fluctuation of the received signal strength indication (RSSI) through PDR, and PDR alleviates the problems of accumulated errors and inaccurate step length estimation through Wi-Fi.

[0059] At present, most deep reinforcement learning indoor positioning methods are designed based on the deep Q-Network (DQN) deep reinforcement learning algorithm, using single Wi-Fi or PDR information for positioning. Although these methods have achieved good positioning results, they ignore the effect of multi-data fusion. In addition, these methods use deep networks to fit the Q function. Since deep networks contain a large number of connection parameters, it takes a lot of time to train positioning models using a large amount of pedestrian data. In addition, if the simulation effect of the deep network is not good, researchers need to rebuild the network to improve performance. Therefore, how to provide a method that can effectively locate and improve the efficiency of training models is an urgent problem to be solved.

[0060] An embodiment of the present invention provides a multi-data fusion BQN indoor positioning method, which can be applied to the indoor positioning process.

[0061] See also Figure 1 , is a flow chart of a multi-data fusion BQN indoor positioning method provided by an embodiment of the present invention. Figure 1As shown, the method mainly comprises the following steps:

[0062] Step 1: Use the Wi-Fi received signal strength indication of the pedestrian trajectory and the heading angle estimated by pedestrian dead reckoning as the first Step agent state Input to the width network;

[0063] Step 2: The wide network extracts features from the agent’s state and outputs an action value vector ,according to - Greedy strategy to select actions ;

[0064] Furthermore, the -The expression of the greedy strategy is

[0065] ;

[0066] Among them, argmax represents the function used to find the action value vector The function of the action when the action Q value reaches the maximum value.

[0067] Step 3: Calculate rewards And the experience sample Store in the experience pool and randomly extract experience samples from the experience pool , calculate the action value vector based on the corresponding action output of the width network;

[0068] Based on the above embodiment, step 3 specifically includes:

[0069] Step 3.1, calculate the reward according to the reward function, where the expression of the reward function is

[0070]

[0071] Among them, the agent is in position Execute an action , move to position , Indicates that the agent is at position The target position, dis function is used to calculate the distance between two positions;

[0072] Step 3.2, from the experience pool Randomly select a certain amount of samples , based on which the target value is calculated , where the target value The expression is

[0073]

[0074] in is based on The maximum action output by the target Q network, is the target Q function;

[0075] Step 3.3, calculate The action value vector , where the action value vector The expression is

[0076]

[0077] in, is a vector representing the Q value of each action, G represents all the actions of the agent, It is the Q function, which is used to calculate the Q value of the action;

[0078] Step 3.4, Corresponding to the action The Q value is to update.

[0079] Step 4: Calculate the parameters of the width network using the ridge regression method. If the fitting accuracy is not met, improve the fitting ability of the model through incremental learning to obtain a trained parameter optimization network model.

[0080] Based on the above embodiment, step 4 specifically includes:

[0081] Step 4.1, change the status As input, it is fed into the left feature node of the first layer, where the expression of the left feature node is

[0082] ;

[0083] Among them, each feature group has a weight matrix and the bias matrix , whose coefficients are randomly generated, is a nonlinear function, the output of each feature node Cascaded , expressed as ;

[0084] Step 4.2, Input into the enhancement group to get the enhanced node output ,

[0085] ;

[0086] in, is the weight matrix, is the bias matrix, and They are all randomly generated. is a nonlinear function;

[0087] Step 4.3: Output the feature group And the enhanced group output is spliced ​​to get , expressed as ;

[0088] Step 4.4, using the ridge regression method, calculate The pseudo-inverse of ,

[0089] ;

[0090] in, is the ridge coefficient, Represents the identity matrix.

[0091] Step 4.5: If the fitting error of width learning exceeds the set threshold, an enhancement node is added, and the additional enhancement node generates a new , the expression is , and are the new weight matrix and bias matrix respectively, and then let , calculated according to the pseudo-inverse theory of block matrices ,

[0092]

[0093] in, Expressed as ,

[0094] ;

[0095] Step 4.6, for The increment of , calculates the new parameters of the width network ,

[0096] ;

[0097] Step 4.7, each time step 1 to step 4.6 is repeated for a preset number of times, the Q network parameters are synchronized with the target Q network parameters, and training is performed until convergence to obtain a trained parameter optimization network model.

[0098] In step 5, the real-time collected Wi-Fi received signal strength indication and the heading angle estimated by pedestrian dead reckoning are input into the trained parameter optimization network model, and the action with the highest value is selected from the action vector output by the width network to estimate the pedestrian position.

[0099] Furthermore, the step 5 specifically includes:

[0100] The real-time Wi-Fi received signal strength indication and the heading angle estimated by pedestrian dead reckoning are input into the model. The agent selects an action in the model output according to the highest Q value and performs this action in combination with the moving distance estimated by pedestrian dead reckoning, thereby transferring to the next state. This process is repeated to obtain the agent's trajectory and estimate the pedestrian's path.

[0101] The multi-data fusion BQN indoor positioning method BQN (Broad Q-network, wide Q network) provided in this embodiment uses the Wi-Fi RSSI and PDR estimated heading angle of the pedestrian trajectory as the first Step agent state Input to the wide network; the wide network extracts features of the agent's state and outputs an action value vector , the agent according to - Greedy strategy to select actions ; Calculate rewards And the experience sample Store in the experience pool. Randomly extract experience samples from the experience pool , calculate the action value vector according to the corresponding action output of the target width network; use the ridge regression method to calculate the parameters of the width network. If the fitting accuracy is not met, improve the fitting ability of the model through incremental learning to obtain the trained parameter optimization network model; input the heading angle estimated by the real-time collected Wi-Fi RSSI and PDR into the trained parameter optimization network model, select the action with the highest value from the action vector output by the width network to estimate the pedestrian position, and convert the parameter optimization method based on the deep network into the parameter optimization method based on the width network. Compared with the existing technology, it maintains good positioning performance while reducing the training time of the model.

[0102] The method of the present invention will be further described below in conjunction with a specific embodiment. In the embodiment of the present invention, the original Wi-Fi RSSI and PDR information are used as the first Step agent state Input into the wide network; the wide network extracts features of the agent's state and outputs an action value vector , the agent according to - Greedy strategy to select actions ; Calculate rewards And the experience sample Store to experience pool. From experience pool Randomly select a certain amount of samples . Use these samples to calculate the target value , and its calculation method is shown in the following formula (1).

[0103] (1)

[0104] in is based on The maximum action output by the target Q network, is the target Q function. Then, according to the following formula (2), The action value vector ,

[0105] (2)

[0106] Here, is a vector representing the Q value of each action, G represents all the actions of the agent, is the Q function, which is used to calculate the Q value of the action. Corresponding to the action The Q value is to update.

[0107] In order to optimize the parameters of the Q network, the ridge regression method is used to calculate the parameters. Specifically, the state As input, it is sent to the left feature node of the first layer. These feature nodes are extracted by the following formula (3):

[0108] (3)

[0109] Each feature group has a weight matrix and the bias matrix , whose coefficients are randomly generated. is a nonlinear function. The output of each feature node Cascaded , expressed as .

[0110] Then as shown below, Is input into the enhancement group to get the enhanced node output ,

[0111] (4)

[0112] in is the weight matrix, is the bias matrix, and They are all randomly generated. It is also a nonlinear function.

[0113] Then, the feature group is output And the enhanced group output splicing to get , expressed as .

[0114] In order to calculate the parameters of the Q network , using the ridge regression method, calculate The pseudo-inverse of The formula is shown in equation (5):

[0115] (5)

[0116] in is the ridge coefficient.

[0117] If the fitting error of width learning exceeds the set threshold, incremental learning is needed to solve the problem. The solution is to add enhancement nodes. For example, an additional enhancement node generates a new , the expression is , and are the new weight matrix and bias matrix respectively, and then consider According to the pseudo-inverse theory of block matrices, It can be calculated by the following formula (6):

[0118] (6)

[0119] in Expressed as ,

[0120] (7)

[0121] Finally, for The increment of the width network parameters is calculated by the following formula (8):

[0122] (8)

[0123] Repeat this training process 3 times to synchronize the Q network parameters with the target Q network parameters. Finally, train until convergence to obtain a network model with optimized parameters.

[0124] Build as Figure 2 The wide neural network shown in Figure 1 includes an input layer and an output layer. The input layer consists of two parts. The left half is the feature node, which contains weights, biases, and nonlinear functions, and is used to map the original agent state to the feature space. The right half is the enhancement node, which also contains weights, biases, and nonlinear functions, and is used to enhance the features and output them.

[0125] like Figure 3As shown, the parameter optimization network model is trained using the training data.

[0126] Use the test data to verify the trained parameter optimization network model.

[0127] The following compares the embodiment of the present invention with the Wi-Fi indoor positioning method based on DQN (Deep Q-network) deep reinforcement learning. The statistical results of the time required for model training to convergence are shown in Table 1:

[0128] Table 1

[0129]

[0130] The Cumulative Distribution Function (CDF) result is as follows: Figure 4 As shown in the figure, the statistical results show that this method uses a wide network instead of a deep network to fit the Q function, taking advantage of the simple structure and fast calculation speed of the wide network, significantly reducing the time required for indoor positioning model training to convergence. In addition, this method integrates PDR information, providing good positioning performance compared to the method using single Wi-Fi information for positioning.

[0131] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof.

[0132] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A multi-data fusion BQN indoor positioning method, characterized in that: include: Step 1: Use the Wi-Fi received signal strength indication of the pedestrian trajectory and the heading angle estimated by pedestrian dead reckoning as the first The state of the agent Input to the width network; Step 2: The wide network extracts features from the agent’s state and outputs an action value vector ,according to - Greedy strategy to select actions ; Step 3: Calculate rewards And the experience sample Store in the experience pool and randomly extract experience samples from the experience pool , calculate the action value vector based on the corresponding action output of the width network; Step 4: Calculate the parameters of the width network using the ridge regression method. If the fitting accuracy is not met, improve the fitting ability of the model through incremental learning to obtain a trained parameter optimization network model. In step 5, the real-time collected Wi-Fi received signal strength indication and the heading angle estimated by pedestrian dead reckoning are input into the trained parameter optimization network model, and the action with the highest value is selected from the action vector output by the width network to estimate the pedestrian position.

2. The method according to claim 1, characterized in that: Said -The expression of the greedy strategy is ; in, Represents the vector used to find the action value The function of the action when the action Q value reaches the maximum value.

3. The method according to claim 2, characterized in that The step 3 specifically includes: Step 3.1, calculate the reward according to the reward function, where the expression of the reward function is Among them, the agent is in position Execute an action , move to position , Indicates that the agent is at position The target position, dis function is used to calculate the distance between two positions; Step 3.2, from the experience pool Randomly select a certain amount of samples , based on which the target value is calculated , where the target value The expression is ; in is based on The maximum value action output by the target Q network, is the target Q function; Step 3.3, calculate The action value vector , where the action value vector The expression is in, is a vector representing the Q value of each action, G represents all the actions of the agent, It is the Q function, which is used to calculate the Q value of the action; Step 3.4, Corresponding to the action The Q value is to update.

4. The method according to claim 3, characterized in that The step 4 specifically includes: Step 4.1, change the status As input, it is fed into the left feature node of the first layer, where the expression of the left feature node is ; Among them, each feature group has a weight matrix and the bias matrix , whose coefficients are randomly generated, is a nonlinear function, the output of each feature node Cascaded , expressed as ; Step 4.2, Input into the enhancement group to get the enhanced node output ; ; in, is the weight matrix, is the bias matrix, and They are all randomly generated. is a nonlinear function; Step 4.3: Output the feature group And the enhanced group output is spliced ​​to get , expressed as ; Step 4.4, using the ridge regression method, calculate The pseudo-inverse of , ; in, is the ridge coefficient, represents the identity matrix; Step 4.5: If the fitting error of width learning exceeds the set threshold, an enhancement node is added, and the additional enhancement node generates a new , the expression is , and are the new weight matrix and bias matrix respectively, and then let , calculated according to the pseudo-inverse theory of block matrices , in, Expressed as , ; Step 4.6, for The increment of , calculates the new parameters of the width network , ; Step 4.7, each time step 1 to step 4.6 is repeated for a preset number of times, the Q network parameters are synchronized with the target Q network parameters, and training is performed until convergence to obtain a trained parameter optimization network model.

5. The method according to claim 4, characterized in that The step 5 specifically includes: The real-time Wi-Fi received signal strength indication and the heading angle estimated by pedestrian dead reckoning are input into the model. The agent selects an action in the model output according to the highest Q value and performs this action in combination with the moving distance estimated by pedestrian dead reckoning, thereby transferring to the next state. This process is repeated to obtain the agent's trajectory and estimate the pedestrian's path.