A method and system for constructing and locating radio maps in a wireless communication network.

Abstract

This application discloses a method and system for constructing and locating radio maps in a wireless communication network. The method specifically includes the following steps: collecting radio spectrum data; selecting edge users to participate in training in response to the completion of radio spectrum data collection; performing federated learning training to obtain a trained neural network model in response to the selected edge users; outputting and saving the neural network model; and locating the radio map based on the saved neural network model. This application, by protecting user privacy, uses a neural network to accurately generalize and generate a high-precision radio map that meets the requirements, making it more suitable for intelligent scenarios.

Description

Technical Field

[0001] This application relates to the field of mobile communication networks, and more particularly to the construction and positioning methods and systems for radio maps in wireless communication networks. Background Technology

[0002] With the rapid development of the global economy, modern mobile communication devices are becoming increasingly dense and heterogeneous. Wireless communication network environments are also becoming more complex, making them difficult to effectively represent in real-world physical environments. Therefore, obtaining information on wireless network coverage and channel quality in real-world scenarios is challenging, posing a significant challenge to the effective management of wireless communication networks in the future. Thus, efficiently and accurately constructing radio maps of known areas is of great significance to the development of future wireless communication networks.

[0003] However, constructing accurate radio maps requires a large amount of radio spectrum data, and the methods for acquiring this data are very limited. Existing methods involve centrally processing radio spectrum data from multiple data owners to generalize a global radio map. However, once data is shared, there is a risk of data leakage, posing a significant challenge to constructing accurate radio maps while protecting data security. Against this backdrop, Federated Learning (FL) helps to accurately train a global network model while protecting user data privacy. Federated Learning ensures that data does not leave the local machine; training participants only exchange model parameters trained on their local networks. It also supports multiple data owners participating in training, accelerating model training. Furthermore, radio maps... Figure 1 However, once accurately constructed, it can be directly applied to real-world scenarios. One important application is for users to use known radio maps to accurately locate themselves and find paths with good channel conditions to complete tasks.

[0004] While federated learning offers an effective solution for protecting data privacy and constructing global radio map network models with near-lossless data transfer, it still faces many challenges, including:

[0005] 1) During federated learning training of radio map models, multiple exchanges of model parameters are required, consuming significant communication resources. Due to limited communication resource blocks, reasonable allocation of these resources is necessary to optimize the entire training process. 2) Local users have limited onboard energy. When radio map model training time is excessively long or the required training accuracy is too high, the onboard energy of some local users may be insufficient to support the entire training process. Therefore, limited onboard energy is a significant factor affecting training quality during federated learning model training. Thus, selecting high-quality users to participate in federated learning training and reducing overall training time is crucial to conserve users' local energy. 3) During federated learning training of radio map models, model transmission during uplink and downlink communication, as well as local training, introduce communication latency issues. This increases the overall training time and leads to untimely model parameter updates.

[0006] Therefore, how to achieve rational resource allocation, low energy consumption, high accuracy, and low latency in the construction of radio maps through federated learning has become an urgent problem to be solved in this field. Summary of the Invention

[0007] The purpose of this application is to provide a method for joint resource allocation of mobile base station computing and caching that combines IoT slicing services. Under the premise that caching resources occupy a major position, the occupancy of other physical resources is balanced. This algorithm can not only achieve better performance, but also flexibly allocate various physical resources in a balanced manner. It is of great significance to better realize the various strict communication requirements of mobile IoT and provide users with better communication efficiency under the service-oriented characteristics.

[0008] To achieve the above objectives, this invention proposes a method for constructing and locating radio maps in a wireless communication network, specifically including the following steps: collecting radio spectrum data; in response to completing the collection of radio spectrum data, selecting edge users to participate in training; in response to selecting edge users to participate in training, performing federated learning training to obtain a trained neural network model; outputting and saving the neural network model; and locating the radio map based on the saved neural network model.

[0009] As mentioned above, edge users participating in the overall training collect radio spectrum information locally through system equipment. This mainly includes location information and channel-related information, such as the signal-to-interference-plus-noise ratio (SINR) and large-scale channel gain information corresponding to the geographical location.

[0010] As described above, the selection of edge users to participate in training specifically includes the following sub-steps: initializing parameters; determining the probability of all edge users participating in this round of training and all edge users not participating in this round of training being selected in the next round of training based on the initialized parameters; and determining the edge users to participate in the next round of training based on the probability of all edge users participating in this round of training and all edge users not participating in this round of training being selected in the next round of training.

[0011] As shown above, the initialization parameters include the initial maximum number of user selections N, and the number of users l that participate in the entire training process. * N-1 marginal users are randomly selected from {1,2,L,L} users to participate in this round of training.

[0012] As described above, determining the probability of each edge user being selected based on the initialized parameters specifically includes the following sub-steps: training batch data according to the initialized parameters to obtain the gradient change information of edge users participating in this round of training before and after training; edge users participating in this round of training sending the obtained gradient change information of edge users participating in this round of training before and after training to the drone aggregation terminal through uplink data transmission; the drone aggregation terminal using a neural network to predict the gradient change information of users not participating in this round of training using the received gradient change information before and after training; and obtaining the probability of each edge user being selected in the next round of training based on the gradient change information of edge users not participating in this round of training.

[0013] As shown above, the gradient change information before and after training includes the gradient change information ||e| generated when edge user l performs the μ-th training round. lμ || and users who participated in the entire training process * gradient change information

[0014] As shown above, the probability of each edge user being selected in the next round of training is:

[0015]

[0016] Among them l * l' represents users who have been participating in the training process, l' represents users who have not participated in this round of training, N represents the maximum number of users allowed to participate in training, and ||e l'μ || represents gradient change information for marginal users who did not participate in this round of training, ||e lμ || represents the gradient transformation information before and after training.

[0017] As shown above, the total loss function in the federated learning process... Where f(w,x) li ,y liLet w be the loss function for edge user l, where w is the model parameter of the neural network for edge user l, and x is the loss function for edge user l. li Let y be the input vector of edge user l's data. li Edge user l is the output vector of the neural network, and m is the size of each batch of training data for edge user l.

[0018] As mentioned above, the process of conducting federated learning training includes obtaining the transmission rate;

[0019] The transmission rate is specifically expressed as:

[0020]

[0021] Where b lμ The bandwidth allocated to user l in the μth round of training, h lμ p represents the channel gain between the UAV and edge user l during the μth round of training. l N0 is the transmit power allocated to user l for uplink communication or downlink communication of UAV, where N0 is the noise power spectral density.

[0022] A system for constructing and locating a radio map in a wireless communication network includes a collection unit, a selection unit, a training unit, an output unit, and a positioning unit. The collection unit collects radio spectrum data. The selection unit selects users to participate in training. The training unit performs federated learning training in response to the selected users to obtain a trained neural network model. The output unit outputs and stores the neural network model. The positioning unit performs radio map positioning based on the stored neural network model.

[0023] This application has the following beneficial effects:

[0024] (1) This invention further accelerates the construction speed of radio maps by optimizing user selection, resource allocation and transmission mechanism in the training process, saving the onboard energy of edge users and reducing communication latency, thereby reducing the cost of building radio maps.

[0025] (2) This application uses a method that protects user privacy and uses a neural network to accurately generalize a high-precision radio map that can meet the needs, making it more suitable for intelligent scenarios.

[0026] (3) This application addresses the method of intelligently constructing radio maps, which can reasonably coordinate the accuracy requirements of a given map with the resource situation of edge users participating in training, and realize the construction of tasks in real-world scenarios.

[0027] (4) In this application, users can download radio maps online or offline from the APP or cloud for applications such as positioning and pre-design of task trajectories, thereby saving the consumption of real resources. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0029] Figure 1 This is a flowchart of a method for constructing and locating a radio map in a wireless communication network according to an embodiment of this application;

[0030] Figure 2 This is a schematic diagram of a system structure for high-precision radio map construction and positioning provided according to an embodiment of this application. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0032] This invention proposes a method and system for constructing and locating radio maps in wireless communication networks. This allows multiple edge data owners to build accurate radio maps without leaving their local networks, solving the problem of data silos and accelerating the construction speed of radio maps. Furthermore, it reduces construction costs, saves energy for edge users, and minimizes communication latency during the radio map construction process. End users can download the generated radio maps online or offline for precise positioning, thus enabling various real-world applications of radio maps.

[0033] This application first addresses the task of collaboratively constructing a global radio map when multiple edge user radio data owners cannot share data. Then, based on the constructed high-precision radio map, precise positioning is performed to complete each user's task. The federated learning-based radio map construction ensures that edge user data does not leave their local machine and can construct a high-precision radio map with almost no loss, protecting data privacy. Furthermore, algorithms to accelerate radio map construction and save energy are proposed to optimize the radio map construction process.

[0034] Scenario Assumptions: Assume that during the training and generation of the radio map through a neural network, a total of L edge users can participate in the federated learning training. Due to the limitation of radio resource blocks, at most N ≤ L users can participate in the overall training task at any given time. The required radio map accuracy is γ. Let u represent the location deployment matrix of the edge users.

[0035] like Figure 1 The diagram illustrates a method for constructing and locating a radio map in a wireless communication network, as provided in this application. The method specifically includes the following steps:

[0036] Step S110: Collect radio spectrum data.

[0037] Before the system is put into operation, each participant in the training first collects radio spectrum information for their respective areas.

[0038] Specifically, edge users participating in the overall training collect radio spectrum information locally through system equipment. This mainly includes location information and channel-related information, such as the signal-to-interference-plus-noise ratio and large-scale channel gain corresponding to geographical location information.

[0039] Furthermore, the received radio spectrum information R at a certain location is represented as R = {q(x,y,H),Q}, where q(x,y,H) in R is the location information of a certain area where the edge user is located, x is the horizontal coordinate mapped to the horizontal plane, y is the vertical coordinate mapped to the horizontal plane, H is the height of the location, and Q is the radio channel information at that location, which can be the signal-to-interference-plus-noise ratio, large-scale channel gain, or small-scale channel gain, etc.

[0040] Step S120: In response to the completion of radio spectrum data collection, select users to participate in the training.

[0041] After the radio spectrum information collection is completed, the federated learning aggregation processing unit uses a neural network to select users to participate in the training. If edge users can accelerate the construction of the radio map, they are selected multiple times to join the training process; otherwise, only a small number of users are selected.

[0042] Specifically, the edge user stores the collected information on the server, and then the drone aggregation terminal uses the gradient change information ||e generated by the edge user l during the μ-th round of training. lμ This allows for more informed user selection, thereby accelerating the overall training process. lμ || represents the magnitude of gradient information change before and after each round of training for edge users.

[0043] Because drones need to utilize gradient information from edge users before and after training ||e lμThe drone needs to obtain gradient change information for all edge users to select users for the next round of training. For easy differentiation, ||e is used. lμ ||For gradient change information of users participating in this round of training,||e l'μ ||This represents gradient change information for users who did not participate in this round of training. Based on ||e lμ ||and||e l'μ The value of || indicates the probability of each user being selected in the next round.

[0044] Step S120 specifically includes the following sub-steps:

[0045] Step S1201: Initialize parameters.

[0046] Among these, the maximum number of user selections N is initialized, and the number of users l who participate in the entire training process is determined. * N-1 marginal users are randomly selected from {1,2...L} users to participate in this round of training.

[0047] Step S1202: Based on the initialized parameters, determine the probability of all edge users participating in this round of training and all edge users not participating in this round of training being selected in the next round of training.

[0048] Step S1202 specifically includes the following sub-steps:

[0049] Step S12021: Train the batch data according to the initialization parameters to obtain the gradient transformation information of the edge users participating in this round of training before and after training.

[0050] The gradient change information before and after training includes the gradient change information ||e| generated when edge user l performs the μ-th training round. lμ || and users who participated in the entire training process * gradient change information

[0051] Specifically, the selected local edge users are trained using batch data based on collected wireless spectrum information, and then the gradient change information ||e| of edge user l before and after the μth training round is obtained. lμ || and users who participated in the entire training process * gradient change information

[0052] When edge users train neural network models locally, they use gradient descent. Assume an edge user trains on J sets of data each time, with each set of data (x...)... j ,y j (Including input x) j With output y jGiven a loss function f(g) and a learning rate λ, the update formula for the model parameters α of the neural network from the μth training round to the μ+1th training round is expressed as:

[0053]

[0054] Where a μ+1 Let a represent the parameters of the neural network model during the (μ+1)th training round. μ This represents the parameters of the neural network model during the μ-th training round. This represents the gradient of the loss function for edge users.

[0055] Using the above formula, we can obtain the gradient changes in the model parameters of the marginal user before and after the μth training round ||e lμ || and users who participated in the entire training process * gradient change information

[0056] Step S12022: The edge users participating in this round of training send the gradient change information of the edge users before and after training to the drone aggregation terminal through uplink data transmission.

[0057] Step S12023: The drone aggregation terminal uses a neural network to utilize the gradient change information received from edge users participating in this round of training ||e lμ ||With users who participated in the entire training process * gradient change information Predict gradient changes for users who did not participate in this round of training ||e l'μ ||.

[0058] Specifically, the formula obtained above... Will The input is passed through the input layer of the neural network, where This represents the data that the neural network needs to train, which includes the ID l of the l-th user and the number l of the user who participated in the entire training process. * gradient change information The marginal users who participated in this round of training ||e lμ || Used as labels for training neural networks.

[0059] After training the neural network, any edge user l will input the vector. The input vector is fed into the neural network through its input layer, and then the neurons in the hidden layers learn the input vector. The nonlinear relationship with the output vector o allows us to output the ||e| of users who did not participate in this round of training. l'μ The following describes how the hidden layer learns the input vector through its neurons. The nonlinear relationship between the output vector o and the output vector o.

[0060] The state θ of the neurons in the hidden layer is:

[0061]

[0062] Where ν in This is the weight matrix of the input vector and the connection strength of neurons in the hidden layer. For users who participate in the entire federal learning and training process * The weight matrix, b θ For the bias vector, the function Let g be the activation function of the neural network, where exp(g) is the exponential function.

[0063] Given the state θ of a neuron, the output vector o can be derived:

[0064] o = v out θ+b o

[0065] Where v out Let b be the weight matrix of the output vector and the connection strength of the neurons in the hidden layer. o This is the bias vector.

[0066] Step S12024: Based on the gradient change information of marginal users who did not participate in this round of training ||e l'μ ||Gradient change information relative to edge users participating in this round of training||e lμ ||, to obtain the probability of each edge user being selected in the next round of training.

[0067] For marginal users who did not participate in this round of training, ||e l'μ The drone aggregation platform uses a neural network for prediction, aiming to accurately predict gradient changes in edge users not involved in the training. l'μ ||, Select a user l * Participated in the entire training process. Then, based on ||e lμ ||and||e l'μ The value of || can be used to calculate the probability of each marginal user being selected in the next round of training, i.e.:

[0068]

[0069] Among them l * l' represents users who have been participating in the training process, l' represents users who have not participated in this round of training, and N represents the maximum number of users allowed to participate in training.

[0070] Step S1203: Determine the edge users to participate in the next round of training based on the probability of all edge users who participated in this round of training being selected in the next round of training compared to all edge users who did not participate in this round of training.

[0071] Specifically, the probability range for each edge user to be selected is set according to the probability magnitude of the edge users. First, the probability of each edge user among the L edge users participating in the current training round in the μth round is calculated using the formula in S12024, assuming it to be P. 1μ P 2μ ... P Lμ Then the probability interval for these L marginal users can be obtained as (0, P). 1μ ), (P 1μ ,P 1μ +P 2μ ), (P 1μ +P 2μ ,P 1μ +P 2μ +P 3μ )……、(P 1μ +P 2μ +K+P L-1μ ,P 1μ +P 2μ +K+P Lμ Then, a random number in the interval (0,1) is generated. If the random number is within the probability interval of the marginal user, the marginal user is selected to participate in the training of the (μ+1)th round (i.e. the next round); otherwise, the marginal user is not selected to participate in the training of the (μ+1)th round.

[0072] This includes, after identifying the edge users who will participate in the next round of training, allocating resources based on the identified edge users who will participate in the current round of training.

[0073] Specifically, based on the specific location of the edge user, the specific location of the drone, the user selection, and the size of the total resource block (such as bandwidth), techniques such as convex optimization (CVX) and interior point method are used to allocate resource blocks reasonably, thereby reducing latency in the communication process.

[0074] Step S130: In response to selecting edge users to participate in training, federated learning training is performed to obtain a trained neural network model.

[0075] After the users participating in the training and the resource allocation are completed, the edge users selected to participate in this round of training use their local radio spectrum data for federated learning training. The essence of federated learning training is that edge users, without sharing their local data, only exchange model parameters with the aggregation point of the federated learning to jointly train a neural network model.

[0076] Assume there are L edge users, some far apart, participating in federated learning training within an edge node network. Each edge user possesses radio spectrum data for their respective region, and the data from these regions are non-overlapping. Since ground base stations cannot fully cover the areas where edge users communicate, this embodiment uses unmanned aerial vehicles (UAVs) as the aggregation processing endpoint for federated learning. The overall loss function during federated learning is defined in this step as follows:

[0077]

[0078] f(w,x li ,y li Let w be the loss function for edge user l, where w is the model parameter of the neural network for edge user l, and x is the loss function for edge user l. li Let y be the input vector of edge user l's data. li Edge user l is the output vector of the neural network, and m is the size of each batch of training data for edge user l.

[0079] Each edge user has its own local neural network during training.

[0080] After performing the following sub-steps of federated training until the loss function value is less than a specified value, proceed to step S140. The sub-steps include:

[0081] Step S1301: Initialize parameters.

[0082] Specifically, initialize the accuracy γ of the radio map, the global model parameters w0, and the number of local iterations Q per round of training.

[0083] Step S1302: Radio spectrum data collected during edge user training.

[0084] Specifically, edge users train the collected radio spectrum data R = {q(x,y,H),Q} through their local neural network, where the edge user uses the location information q(x,y,H) as the input to the local neural network and the channel information Q as the label for training the local neural network to train the data.

[0085] Edge users participating in the training randomly select R sets of data each time for Q iterations of training.

[0086] Step S1303: In response to the completion of training the radio spectrum data, update the parameters of the neural network model.

[0087] When edge users train locally, they use gradient descent to update the parameters of the neural network model.

[0088]

[0089] Edge users use this formula to reduce the loss function in S130 until the loss function is reduced to a specified threshold.

[0090] in For the gradient of the loss function for edge users, w lμ Let w be the neural network model parameters for edge user l during the μth round of training. l,μ+1 Let be the model parameters for edge user l during the (μ+1)th training round, and k be the learning rate for edge user l.

[0091] Step S1303: Edge users upload updated neural network model parameter information.

[0092] The edge users participating in the training will use the model parameters w from the μth training round. lμ The data is transmitted to the drone aggregation terminal via uplink communication.

[0093] The process of transmitting model parameters also includes obtaining the transmission rate. The transmission rate represents how quickly the drone and the edge user exchange model parameters; a higher value indicates faster data transmission between the drone and the edge user, and vice versa.

[0094] The smaller the value, the greater the delay in model parameter transmission during communication, which leads to increased communication latency and increases the time required to build the radio map. Assuming the size of the model parameters exchanged between the drone and the local user in each round is W bits, and the system's allowable transmission delay is t, then the transmission rate r... lμ r needs to be satisfied lμ t≥W, otherwise the system's transmission delay requirement is not met. Where the transmission rate r... lμ Specifically, it is expressed as follows:

[0095]

[0096] Where b lμ The bandwidth allocated to user l in the μth round of training, h lμ p represents the channel gain between the UAV and edge user l during the μth round of training. l N0 is the transmit power allocated to user l for uplink communication or downlink communication of UAV, where N0 is the noise power spectral density.

[0097] Step S1304: The drone processes the collected model parameters through aggregation and weighting to obtain the initial values ​​w of the neural network model parameters for the next round of training for edge users. μ+1 and through downlink communication w μ+1 It is transmitted to various peripheral users.

[0098] Where wμ+1 Specifically, it is expressed as follows:

[0099]

[0100] Where M is the data size of all edge users participating in the training.

[0101] w is transmitted via downlink communication. μ+1 The process of transmitting data to various edge users also includes obtaining the transmission rate.

[0102] Step S1305: Each edge user assigns the received neural network model parameters to its own neural network.

[0103] The edge user will receive the model parameters w μ+1 Assign values ​​to the local neural networks of each edge user, and repeat steps S1302-S1304 until the neural network reaches the maximum number of iterations Q or the neural network model reaches the specified accuracy.

[0104] Specifically, based on the accuracy set for the radio map, it is determined whether the trained neural network model has met the accuracy requirements of the radio map. Assuming that when MSE ≤ γ (where γ is a given threshold and its value is constant), the trained neural network has achieved the ability to generalize to a radio map of the given accuracy, then the federated learning training is complete, and step S140 is executed. Otherwise, training continues, and steps S120-S130 are repeated until the accuracy requirements of the given radio map are met.

[0105] The MSE is calculated as follows:

[0106]

[0107] Where S is the number of selected predicted radio spectrum data points, and y is the actual radio spectrum data. This refers to radio spectrum data predicted by a neural network.

[0108] Step S140: Output and save the neural network model.

[0109] Specifically, the neural network model of the final radio map that meets the generalization requirements is saved to the APP or the cloud, and users can download it online or offline and apply it to various real-world scenarios.

[0110] Step S150: Perform radio map positioning based on the saved neural network model.

[0111] Step S150 specifically includes the following sub-steps:

[0112] Step S1501: The user downloads the radio map.

[0113] Users can download neural network models that can accurately generalize radio maps from the app or cloud.

[0114] Step S1502: The user inputs the radio spectrum value of the target to be located into the neural network model.

[0115] Step S1503: The neural network outputs the precise location of the target based on the neural network model.

[0116] Step S1504: The user obtains the location information of the target.

[0117] Example 2

[0118] This application provides a system for constructing and locating radio maps in a wireless communication network, specifically including a collection unit 210, a selection unit 220, a training unit 230, an output unit 240, and a positioning unit 250.

[0119] The collection unit 210 is used to collect radio spectrum data.

[0120] Selection unit 220 is connected to collection unit 210 and is used to select users to participate in training.

[0121] The training unit 230 is connected to the selection unit 220 and is used to perform federated learning training in response to the selection of users to participate in the training, so as to obtain a trained neural network model.

[0122] The output unit 240 is connected to the training unit 230 and is used to output and store the neural network model.

[0123] The positioning unit 250 is connected to the output unit 240 and is used to perform radio map positioning based on the stored neural network model.

[0124] This application has the following beneficial effects:

[0125] (1) This invention further accelerates the construction speed of radio maps by optimizing user selection, resource allocation and transmission mechanism in the training process, saving the onboard energy of edge users and reducing communication latency, thereby reducing the cost of building radio maps.

[0126] (2) This application uses a method that protects user privacy and uses a neural network to accurately generalize a high-precision radio map that can meet the needs, making it more suitable for intelligent scenarios.

[0127] (3) This application addresses the method of intelligently constructing radio maps, which can reasonably coordinate the accuracy requirements of a given map with the resource situation of edge users participating in training, and realize the construction of tasks in real-world scenarios.

[0128] (4) In this application, users can download radio maps online or offline from the APP or cloud for applications such as positioning and pre-design of task trajectories, thereby saving the consumption of real resources.

[0129] Although the examples referenced in this application are described for illustrative purposes only and not for limiting the scope of this application, changes, additions and / or deletions to the implementation may be made without departing from the scope of this application.

[0130] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing and locating radio maps in a wireless communication network, characterized in that, Specifically, the following steps are included: To collect radio spectrum data; In response to the completion of radio spectrum data collection, edge users are selected to participate in the training. In response to the selection of edge users to participate in training, federated learning training is performed to obtain a trained neural network model; The output saves the neural network model; Radio map positioning is performed based on the saved neural network model; The selection of edge users to participate in training includes the following sub-steps: Initialization parameters; initialization parameters include initializing the maximum number of user selections N, and determining the number of users l to participate in the entire training process. * N-1 marginal users are randomly selected from {1,2,L,L} users to participate in this round of training; Based on the initialized parameters, determine the probability of all edge users who participated in this round of training and all edge users who did not participate in this round of training being selected in the next round of training; The edge users who will participate in the next round of training are determined based on the probability of all edge users who participated in this round of training being selected in the next round of training compared to all edge users who did not participate in this round of training. The process of determining the probability of each marginal user being selected based on the initialized parameters includes the following sub-steps: Training is performed on batch data based on initialization parameters to obtain gradient transformation information of edge users participating in this round of training before and after training; The edge users participating in this round of training send the gradient change information of the edge users before and after training to the drone aggregation terminal through uplink data transmission; The drone aggregation terminal uses a neural network to predict the gradient change information of users who did not participate in the current training round by utilizing the received gradient change information before and after training. Based on the gradient change information of edge users who did not participate in this round of training, the probability of each edge user being selected for the next round of training is obtained.

2. The method for constructing and locating radio maps in a wireless communication network as described in claim 1, characterized in that, Edge users participating in the overall training collect radio spectrum information locally through system equipment. This mainly includes location information and channel-related information, such as the signal-to-interference-plus-noise ratio (SINR) and large-scale channel gain information corresponding to the geographical location.

3. The method for constructing and locating radio maps in a wireless communication network as described in claim 1, characterized in that, The gradient change information before and after training includes the gradient change information generated by edge user l during the μ-th training round ||e lμ || and users who participated in the entire training process * gradient change information 4. The method for constructing and locating radio maps in a wireless communication network as described in claim 3, characterized in that, The probability P of each marginal user being selected in the next round of training lμ+1 Specifically, it is expressed as follows: Among them l * l' represents users who have been participating in the training process, l' represents users who have not participated in this round of training, N represents the maximum number of users allowed to participate in training, and ||e l'μ || represents gradient change information for marginal users who did not participate in this round of training, ||e lμ || represents the gradient transformation information before and after training.

5. The method for constructing and locating radio maps in a wireless communication network as described in claim 4, characterized in that, Define the overall loss function in the federated learning process. Where f(w,x) li ,y li Let w be the loss function for edge user l, where w is the model parameter of the neural network for edge user l, and x is the loss function for edge user l. li Let y be the input vector of edge user l's data. li Edge user l is the output vector of the neural network, and m is the size of each batch of training data for edge user l.

6. The method for constructing and locating a radio map in a wireless communication network as described in claim 5, characterized in that, The process of conducting federated learning training includes obtaining the transmission rate; Transmission rate r lμ Specifically, it is expressed as follows: Where b lμ The bandwidth allocated to user l in the μth round of training, h lμ p represents the channel gain between the UAV and edge user l during the μth round of training. l N0 is the transmit power allocated to user l for uplink communication or downlink communication of UAV, where N0 is the noise power spectral density.

7. A system for constructing and locating radio maps in a wireless communication network, characterized in that, Specifically, it includes a collection unit, a selection unit, a training unit, an output unit, and a localization unit; The collection unit is used to collect radio spectrum data; The selection unit is used to select edge users to participate in training. The training unit is used to perform federated learning training in response to the selection of users to participate in the training, and to obtain a trained neural network model. The output unit is used to output and store the neural network model; The positioning unit is used for radio map positioning based on the stored neural network model; The selection of edge users to participate in training specifically includes the following sub-steps: Initialization parameters; initialization parameters include initializing the maximum number of user selections N, and determining the number of users l to participate in the entire training process. * N-1 marginal users are randomly selected from {1,2,L,L} users to participate in this round of training; Based on the initialized parameters, determine the probability of all edge users who participated in this round of training and all edge users who did not participate in this round of training being selected in the next round of training; The edge users who will participate in the next round of training are determined based on the probability of all edge users who participated in this round of training being selected in the next round of training compared to all edge users who did not participate in this round of training. The process of determining the probability of each marginal user being selected based on the initialized parameters includes the following sub-steps: Training is performed on batch data based on initialization parameters to obtain gradient transformation information of edge users participating in this round of training before and after training; The edge users participating in this round of training send the gradient change information of the edge users before and after training to the drone aggregation terminal through uplink data transmission; The drone aggregation terminal uses a neural network to predict the gradient change information of users who did not participate in the current training round by utilizing the received gradient change information before and after training. Based on the gradient change information of edge users who did not participate in this round of training, the probability of each edge user being selected for the next round of training is obtained.

Images