Radio map construction method, system, computer device and storable medium

Abstract

The application discloses a radio map construction method and system based on federated learning, computer equipment and a storable medium, the method comprising: designing a radio signal strength prediction neural network model according to the characteristics of the radio map; obtaining parameters such as the number of user dispatches and the number of training rounds, and initializing the model; randomly dispatching users according to the number of user dispatches per round; compressing the model by the server and sending it to the selected users; the selected users locally train and compress the model parameter change amount based on the received model parameters, and upload the compressed parameter change amount to the server; the server aggregates the received parameter change amount, and updates the model parameters based on the aggregated parameter change amount; according to the number of training rounds, repeat the above steps to obtain the final neural network model and construct the radio map.

Description

Technical Field

[0001] This invention relates to the field of radio map construction technology, and specifically to a radio map construction method, system, computer device, and storage medium based on federated learning. Background Technology

[0002] With the rapid development of artificial intelligence technology, the construction of data-driven radio maps requires a large amount of terminal data for training. Therefore, federated learning has enormous application potential in the field of radio map construction. Federated learning is a distributed learning framework. In the federated learning process, a central server broadcasts a global model, and then user-end models are trained on local datasets. The updated model parameters are then sent back to the server, which aggregates the data and repeats this process until the model converges. A major challenge of federated learning is the communication overhead problem. The large amount of parameter transmission and transmission during federated learning incurs significant communication overhead. Therefore, data compression before uplink and downlink transmission is necessary in federated learning, which can greatly reduce the communication overhead required to train the radio signal strength prediction network. Existing federated learning-based radio map construction methods and systems do not simultaneously consider uplink and downlink communication compression. Unrestricted use of compression algorithms can impair model performance and reduce the accuracy of radio signal strength prediction. Therefore, algorithm design is needed to reduce communication overhead while minimizing the impact on the performance of the radio map construction task. Summary of the Invention

[0003] The purpose of this invention is to address the aforementioned deficiencies in the prior art by providing a radio map construction method, system, computer device, and storage medium based on federated learning. This radio map construction method achieves radio signal strength prediction by designing a neural network model, thereby constructing a radio map for a specific area. Furthermore, by designing uplink and downlink transmission compression algorithms and model aggregation algorithms, it reduces communication overhead while mitigating the performance impact of compression on the neural network model.

[0004] The first objective of this invention can be achieved by adopting the following technical solution:

[0005] A method for constructing radio maps based on federated learning, the method comprising:

[0006] Based on the characteristics of radio maps, a neural network model for predicting radio signal strength was designed.

[0007] Obtain parameters including the number of user scheduling and the number of training rounds, and initialize the neural network model;

[0008] Users are randomly assigned based on the number of users scheduled in each round, and users are selected to participate in the federated learning in this round.

[0009] The server compresses the neural network model and sends the compressed neural network model to the selected user.

[0010] The selected users train the model locally based on the model parameters sent by the server;

[0011] The selected user compresses the parameter changes of the locally trained model and uploads the compressed parameter changes to the server.

[0012] The server aggregates the parameter changes uploaded by the selected users and updates the model parameters based on the aggregated parameter changes.

[0013] Based on the number of training rounds, the steps of randomly scheduling users, downlink compressed transmission, user local training, uplink compressed transmission, and model aggregation and update are repeated to complete the training of the neural network model.

[0014] A neural network model is used to predict the radio signal strength of a specific area, resulting in a radio map.

[0015] Furthermore, in the design of the neural network model for predicting radio signal strength based on the characteristics of radio maps,

[0016] The radio map is characterized by the dimensionality of the input data, the dimensionality of the output predicted values, and the number of base stations involved in the radio map. Based on these characteristics, the number of neurons in the input and output layers of the neural network model can be determined.

[0017] The neural network model includes a shared network and a personalized network;

[0018] The shared network is used to predict the radio signal strength of all base stations;

[0019] The number of personalized networks is consistent with the number of base stations involved in the radio map, and they are only responsible for predicting the radio signal strength of the corresponding base station. The joint design of shared networks and personalized networks is conducive to the neural network model predicting the radio signal strength of each base station involved in the radio map relatively independently.

[0020] The input data passes through the shared network and the personalized network in sequence. The personalized network outputs the predicted value of the radio signal strength of the corresponding base station.

[0021] Furthermore, the process of acquiring relevant parameters, including the number of user scheduling attempts and the number of training rounds, and initializing the neural network model,

[0022] The number of users scheduled is the number of users selected to participate in federated learning each round.

[0023] The number of training rounds is the total number of times that the user is repeatedly randomly scheduled, downlink compressed transmission is performed, user local training is conducted, uplink compressed transmission is performed, and model aggregation updates are performed.

[0024] The relevant parameters also include learning rate, data batch size, number of local iterations, uplink compression quantization bits, uplink compression scaling factor, uplink compression error feedback factor, downlink compression quantization bits, and momentum coefficient of model aggregation.

[0025] The initialization of the neural network model involves randomly assigning initial values ​​to the model parameters.

[0026] Furthermore, the process of randomly scheduling users based on the number of users scheduled in each round, and selecting users to participate in the current round of federated learning,

[0027] Based on the number of users to be scheduled and the random scheduling algorithm used, users to be scheduled are randomly selected from all available users for this round of training.

[0028] Furthermore, the server compresses the model and sends the compressed model to the selected user as follows:

[0029] The number of downlink compressed quantization levels is determined based on the downlink compressed quantization bit depth.

[0030] Based on the number of quantization levels and the maximum and minimum values ​​of the server's model parameters to be sent, the server's model parameters to be sent are subjected to min-max normalization. This normalizes the range of absolute values ​​of the model parameters to be sent to be greater than or equal to zero and less than or equal to the number of downlink compressed quantization levels minus one, resulting in a normalized model parameter vector. This normalization step is a prerequisite for further quantization encoding of the model parameter vector.

[0031] The normalized model parameter vector is randomly rounded to obtain the parameters to be sent. The server then quantizes and encodes the model parameter vector to be sent and sends it to the selected user. After quantization and encoding, the total number of bits in the model parameter vector to be sent decreases, reducing downlink communication overhead.

[0032] The parameters that have already been issued are denormalized to obtain the quantized model parameters;

[0033] The model parameters on the server are updated to the quantized model parameters.

[0034] Furthermore, the selected user trains the neural network model locally based on the learning rate, data batch size, and local iteration number issued by the server, and the neural network model is updated locally towards the local optimum.

[0035] Furthermore, the process by which the selected user compresses the parameter changes of the locally trained model and uploads the compressed parameter changes to the server is as follows:

[0036] Calculate the change in model parameters based on the model parameters before and after local training.

[0037] The number of unscheduled rounds is calculated based on the difference between the number of rounds the selected user had in the current round and the number of rounds the user was selected in the previous round. The uplink compression error feedback factor is then multiplied by the number of unscheduled rounds raised to the power of the error feedback coefficient.

[0038] The vector to be compressed is obtained by multiplying the change in model parameters by the compression residual vector generated during the previous uplink compression and the error feedback coefficient. This error feedback mechanism can reduce the error in the change in model parameters caused by the accumulation of residuals.

[0039] The number of uplink compressed quantization levels is determined based on the uplink compressed quantization bit depth.

[0040] The vector to be compressed is multiplied by the compression scaling factor of the previous row and scaled. Elements whose absolute value after scaling is greater than the number of quantization levels minus one are limited to the number of quantization levels minus one.

[0041] If the maximum absolute value after scaling is less than the number of quantization levels minus one, then multiply it by a coefficient so that its maximum absolute value is equal to the number of quantization levels minus one, and obtain the normalized vector.

[0042] The normalized vector is subjected to random rounding and denormalization in sequence to obtain the compressed model parameter change after error feedback to be uploaded by the user, and the change is uploaded to the server. The total number of bits in the compressed model parameter change is lower than that in the original model parameter change, and the uplink communication overhead is effectively reduced.

[0043] Calculate the difference between the vector to be compressed and the model parameter change after the error feedback to be uploaded, obtain the compression residual vector generated by the row compression in this round, and record it locally on the user's device.

[0044] Furthermore, the server aggregates the parameter changes uploaded by the selected users and updates the model parameters based on the aggregated parameter changes as follows:

[0045] The server averages the changes in model parameters uploaded by users in the current round to obtain the aggregate vector for that round. The aggregate vector represents the basic update direction of the neural network model, and the accuracy of the neural network model in predicting signal strength gradually improves with iterative model updates.

[0046] The model update amount for this round is obtained by adding the aggregation vector of the previous round to the model update amount of the previous round multiplied by the momentum coefficient of the model aggregation. The aggregation algorithm with momentum term can speed up the convergence speed of model training.

[0047] The server's model parameters are added to the model update amount in this round to obtain the aggregated model parameters.

[0048] The second objective of this invention can be achieved by adopting the following technical solution:

[0049] A federated learning-based radio map building system, the radio map building system comprising:

[0050] The neural model design module is used to design a neural network model for predicting radio signal strength based on the characteristics of radio maps.

[0051] The model initialization module is used to obtain parameters including the number of user scheduling and the number of training rounds, and to initialize the neural network model.

[0052] The random user scheduling module is used to randomly schedule users based on the number of users scheduled in each round, and select users to participate in the current round of federated learning.

[0053] The downlink compression transmission module is used by the server to compress the neural network model and send the compressed neural network model to the selected user;

[0054] The user local training module is used by selected users to train the model locally based on the model parameters sent by the server.

[0055] The uplink compression transmission module is used by the selected user to compress the parameter changes of the locally trained model and upload the compressed parameter changes to the server.

[0056] The model aggregation and update module is used by the server to aggregate the parameter changes uploaded by the selected users and update the model parameters based on the aggregated parameter changes.

[0057] The neural model training module repeats the steps of random user scheduling, downlink compressed transmission, user local training, uplink compressed transmission, and model aggregation update according to the number of training rounds to complete the training of the neural network model.

[0058] The radio map generation module calls a neural network model to predict the radio signal strength of a specific area and obtain a radio map.

[0059] The third objective of this invention can be achieved by adopting the following technical solution:

[0060] A computer device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the above-described radio map construction method.

[0061] The fourth objective of this invention can be achieved by adopting the following technical solution:

[0062] A storage medium storing a program that, when executed by a processor, implements the above-described radio map construction method.

[0063] The present invention has the following advantages and effects compared with the prior art:

[0064] (1) This invention achieves the function of radio signal strength prediction by designing a neural network model, thereby realizing the purpose of constructing a radio map for a specific area. The neural network model in this invention consists of a shared network and a personalized network, ensuring that the neural network model can both extract features from the global radio map and make the radio signal strength prediction of each base station in the radio map relatively independent. This design significantly accelerates the convergence speed of the neural network model and improves the accuracy of radio signal strength prediction.

[0065] (2) This invention designs uplink and downlink transmission compression algorithms for radio map construction based on federated learning. Both uplink and downlink transmission compression algorithms use quantization encoding to reduce the total number of bits transmitted. In the uplink transmission, an error feedback mechanism is also introduced to reduce the error in model parameter changes caused by residual accumulation. This design significantly reduces the communication overhead of radio map construction based on federated learning while minimizing the performance loss of the neural network model due to compression. Attached Figure Description

[0066] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0067] Figure 1 This is a flowchart of the steps of a radio map construction method based on federated learning disclosed in this invention;

[0068] Figure 2 This is a structural block diagram of a radio map construction system based on federated learning disclosed in Embodiment 3 of the present invention;

[0069] Figure 3 This is a schematic diagram illustrating the change in the training loss function of the radio signal prediction network during the process of constructing a radio map using federated learning in Embodiment 1 of the present invention.

[0070] Figure 4 This is a schematic diagram comparing the total communication overhead generated by downlink transmission using compression and not using compression during the process of building a radio map using federated learning in Embodiment 1 of the present invention;

[0071] Figure 5 This is a schematic diagram comparing the total communication overhead generated by using compression and not using compression during the uplink transmission in the process of building a radio map through federated learning in Embodiment 1 of the present invention.

[0072] Figure 6 This is a schematic diagram illustrating the change in the training loss function of the radio signal prediction network during the federated learning process of constructing a radio map in Embodiment 2 of the present invention;

[0073] Figure 7 This is a schematic diagram comparing the total communication overhead generated by downlink transmission using compression and not using compression during the process of building a radio map using federated learning in Embodiment 2 of the present invention;

[0074] Figure 8 This is a schematic diagram comparing the total communication overhead generated by uplink transmission using compression and not using compression during the federated learning process of constructing a radio map in Embodiment 2 of the present invention.

[0075] Figure 9 This is a structural block diagram of the computer device in Embodiment 4 of the present invention. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Example 1

[0078] This embodiment provides a method for constructing radio maps based on federated learning. The following specific example illustrates the implementation of the invention, and simulation results are given to demonstrate the effectiveness of the invention. The relevant settings of the specific embodiment are as follows: the features of the dataset are 2D vectors representing the coordinates of a point on the map; the labels are 4D vectors representing the radio signal strengths received by the point on the map from the four base stations; there are approximately 360,000 data samples in total; the number of users is 90, that is, each user is used for approximately 4,000 data samples, and 5 users are scheduled to participate in training each round. The training epochs were 500; the optimizer used was ADAM; the loss function was Mean Absolute Error (MAE); the learning rate was initially 1e-3, then 1e-5 for the last 50 epochs, with intermediate changes following a Discrete Cosine Annealing strategy, changing every 50 epochs; the data batch size was 128; the number of local iterations was 10; the uplink compression quantization bits were 2 bits for the first 50 epochs, 6 bits for the last 50 epochs, and 3 bits for the remaining epochs; the uplink compression scaling factor was 256; the uplink compression error feedback factor was 0.9; the downlink compression quantization bits were 11 bits; and the momentum coefficient for model aggregation was 0.4. Since the method provided in this invention involves repeating the loop several times, this embodiment only uses one epoch as an example for illustration, while the final result is the result after the complete implementation of this method.

[0079] Reference Figure 1 This invention provides a method for constructing radio maps based on federated learning, which includes the following steps:

[0080] S1. Based on the characteristics of radio maps, design a neural network model for predicting radio signal strength;

[0081] Specifically, the radio map in this embodiment is characterized by an input dimension of 2 and an output dimension of 4, meaning it involves radio signal prediction from 4 base stations, thus the network has 4 personalized networks. Specifically, the parameterized network layers in this embodiment are all fully connected layers, and the ReLU function is used for activation.

[0082] The hidden layer of the shared network is designed as (128,128,128,128,128,128,256,256,256); the hidden layer of each personalized network is designed as (128,128,128,128,128,64). Each personalized network outputs a 1-dimensional predicted value of radio signal strength. Finally, the predicted values ​​output by the four personalized networks are concatenated into a 4-dimensional output.

[0083] S2. Obtain the number of user schedulers, number of training rounds, learning rate, data batch size, number of local iterations, uplink compression quantization bits, uplink compression scaling factor, uplink compression error feedback factor, downlink compression quantization bits, and momentum coefficient of model aggregation.

[0084] Specifically, the number of user schedulers is 5; the number of training rounds is 500; the learning rate is initially 1e-3, and in the last 50 rounds it is 1e-5, with intermediate changes following a discrete cosine annealing strategy, changing once every 50 rounds; the data batch size is 128; the number of local iterations is 10; the uplink compression quantization bits are 2 bits in the first 50 rounds, 6 bits in the last 50 rounds, and 3 bits in the remaining rounds; the uplink compression scaling factor is 256; the uplink compression error feedback factor is 0.9; the downlink compression quantization bits are 11 bits; the momentum coefficient of model aggregation is 0.4; and the random initialization of neural network parameters uses the Kaiming initialization method.

[0085] S3. Randomly schedule users according to the number of users scheduled in each round, and select users to participate in the federated learning in this round.

[0086] Specifically, in this embodiment, users are sampled using a uniform distribution, and 5 users are randomly selected from 90 users to participate in this round of training.

[0087] S4. The server compresses the model and sends the compressed model to the selected user;

[0088] Specifically, the downlink compressed quantization bit depth is 11 bits, and the number of downlink compressed quantization levels is 2. 11-1 = 1024;

[0089] Let w be the model parameters to be sent to the server, where the maximum absolute value is w. max The smallest absolute value is w min The normalization method is as follows: The normalized vector is then randomly rounded to an integer to obtain the parameters to be sent. The server then sends the model parameters to be sent to the selected user.

[0090] The server performs inverse normalization on the issued parameters to obtain quantized model parameters, and the local model parameters on the server are updated to the quantized model parameters.

[0091] S5. The selected user trains the model locally based on the model parameters sent by the server;

[0092] Specifically, the learning rate is set for the current round number; the data batch size is 128; the number of local iterations is 10; the loss function is MAE; the optimizer is ADAM; and the user trains the model locally based on these hyperparameters.

[0093] S6. The selected user compresses the parameter changes of the locally trained model and uploads the compressed parameter changes to the server.

[0094] Specifically, assuming the user's last scheduled round is τ′, the current round is τ, the compressed residual vector is r, and the model parameter change is Δw, then the vector to be compressed is calculated as g = Δw + r. τ-τ′ ;

[0095] The uplink compressed quantization bit depth B is set according to the current round number, and the number of uplink compressed quantization levels is S = 2. B-1 ;

[0096] Specifically, the normalization operation involves multiplying the vector to be compressed by the previous row compression scaling factor, and then limiting the elements whose absolute value after scaling is greater than the number of quantization levels minus one to the number of quantization levels minus one. This normalization operation can be combined as follows: The normalized vector is then randomly rounded to an integer to obtain the vector to be uploaded, which the user then sends to the server.

[0097] The vector to be uploaded is denormalized to obtain a quantized vector. The difference between the quantized vector and the vector to be compressed is calculated to obtain the compression residual vector generated by the uplink compression in this round, and recorded locally by the user.

[0098] S7. The server aggregates the parameter changes uploaded by the selected users and updates the model parameters based on the aggregated parameter changes:

[0099] Specifically, assume the model update amount in the previous round is g. t-1 The average value of the vectors uploaded by users in this round is Given a momentum coefficient of 0.4, the aggregated amount of momentum update can be expressed as follows: Assume the model parameters of the previous round are w. t-1 Then the model parameter update for this round can be represented as w t =w t-1 +g t .

[0100] Reference Figure 3 After 500 rounds of training, the average absolute error of the neural network model for predicting radio signal strength decreased to 0.05, meaning that the average prediction error for radio signal strength on the validation set was 0.05 dBm, demonstrating the effectiveness of the invention. (Refer to...) Figure 4 and Figure 5After 500 rounds of training, the total communication overhead of the scheme without uplink and downlink compression algorithms was 1.57Gb, while the total uplink overhead with compression algorithms was 0.14Gb and the total downlink overhead was 0.42Gb. This shows that the uplink and downlink compression scheme of the present invention can effectively reduce communication overhead.

[0101] Example 2

[0102] This embodiment provides a method for constructing a radio map based on federated learning when the total number of data samples is small. The following are the relevant settings for a specific embodiment: the features of the dataset are 2D vectors representing the coordinates of a point on the map; the labels are 4D vectors representing the radio signal strengths received by the point on the map from four base stations; there are approximately 160,000 data samples in total, about half of that in Embodiment 1; the number of users is 40, meaning each user has approximately 4,000 data samples, and 3 users are scheduled to participate in training each round. Other settings, such as the number of training rounds, optimizer, loss function, learning rate, uplink and downlink compression quantization bits, and model aggregation momentum coefficient, are the same as in Embodiment 1. Since the method provided by this invention involves repeating the loop several times, this embodiment only takes one round as an example for illustration, and the final effect diagram is the result after the complete implementation of this method.

[0103] S1. Based on the characteristics of radio maps, design a neural network model for predicting radio signal strength;

[0104] Specifically, the radio map in this embodiment is characterized by an input dimension of 2 and an output dimension of 4, involving radio signal prediction from 4 base stations. Therefore, the network has 4 personalized networks. Specifically, the parameterized network layers in this embodiment are all fully connected layers, and the activation function used is the LeakyReLU function. Due to the small total data sample size, the hidden layers of the shared network use a shallow design, with each hidden layer having a size of (128, 128, 128, 256, 256). Simultaneously, the hidden layers of each personalized network also use a shallow design, with each hidden layer having a size of (128, 128, 64). Each personalized network outputs a 1-dimensional radio signal strength prediction value. Finally, the prediction values ​​output by the 4 personalized networks are concatenated into a 4-dimensional output. The size of the neural network model in this embodiment is approximately half that of Embodiment 1.

[0105] S2. Obtain the number of user schedulers, number of training rounds, learning rate, data batch size, number of local iterations, uplink compression quantization bits, uplink compression scaling factor, uplink compression error feedback factor, downlink compression quantization bits, and momentum coefficient of model aggregation, and initialize the neural network model.

[0106] Specifically, due to the small total data sample size and the smaller scale of the neural network model, the obtained parameters are slightly different from those in Example 1. The number of user schedulers is 3; the uplink compression quantization bits are 3 bits for the first 50 rounds, 9 bits for the last 50 rounds, and 5 bits for the remaining rounds; the uplink compression scaling factor is 256; the uplink compression error feedback factor is 0.9; the downlink compression quantization bits are 16 bits; the number of training rounds, learning rate, data batch size, number of local iterations, uplink compression scaling factor, uplink compression error feedback factor, and momentum coefficient of model aggregation are all the same as in Example 1; and the random initialization of the neural network parameters uses the Kaiming initialization method.

[0107] S3. Referring to Example 1, users are randomly scheduled according to the number of users scheduled in each round, and users participating in this round of federated learning are selected.

[0108] S4. Referring to Example 1, the server compresses the model and sends the compressed model to the selected user;

[0109] S5. Referring to Example 1, the selected user trains the model locally based on the model parameters sent by the server.

[0110] S6. Referring to Example 1, the selected user compresses the parameter changes of the locally trained model and uploads the compressed parameter changes to the server.

[0111] S7. Referring to Example 1, the server aggregates the parameter changes uploaded by the selected user and updates the model parameters based on the aggregated parameter changes.

[0112] Reference Figure 6 Even though the total number of data samples in this embodiment is about half that in Embodiment 1, after 500 rounds of training, the average absolute error of the neural network model for predicting radio signal strength can be reduced to 0.11, that is, the average prediction error for radio signal strength on the validation set is 0.11 dBm. This demonstrates that the present invention can construct a highly accurate radio map under different total numbers of data samples. (Refer to...) Figure 7 and Figure 8 After 500 rounds of training, because the neural network model in this embodiment is smaller than that in Embodiment 1, the total communication overhead of the scheme without uplink and downlink compression algorithms is 0.39Gb, which is less than the 1.57Gb in Embodiment 1. The total uplink overhead with compression algorithms is 0.045Gb, and the total downlink overhead is 0.22Gb, indicating that the uplink and downlink compression scheme of the present invention can effectively reduce communication overhead.

[0113] Example 3

[0114] like Figure 2As shown, this embodiment provides a radio map construction system based on federated learning. The radio map construction system includes: a neural model design module 201, a model initialization module 202, a random user scheduling module 203, a downlink compressed transmission module 204, a user local training module 205, an uplink compressed transmission module 206, a model aggregation and update module 207, a neural model training module 208, and a radio map generation module 209. The specific functions of each module are as follows:

[0115] The neural model design module 201 is used to design a neural network model for predicting radio signal strength based on the characteristics of radio maps.

[0116] The model initialization module 202 is used to obtain parameters including the number of user scheduling and the number of training rounds, and to initialize the neural network model;

[0117] The random user scheduling module 203 is used to randomly schedule users according to the number of users scheduled in each round, and select users to participate in the current round of federated learning.

[0118] The downlink compression transmission module 204 is used by the server to compress the neural network model and send the compressed neural network model to the selected user;

[0119] The user local training module 205 is used for the selected user to train the model locally based on the model parameters sent by the server.

[0120] The uplink compression transmission module 206 is used by the selected user to compress the parameter changes of the locally trained model and upload the compressed parameter changes to the server.

[0121] The model aggregation and update module 207 is used by the server to aggregate the parameter changes uploaded by the selected user and update the model parameters based on the aggregated parameter changes.

[0122] The neural model training module 208 repeats the steps of random user scheduling, downlink compressed transmission, user local training, uplink compressed transmission, and model aggregation update according to the number of training rounds to complete the training of the neural network model.

[0123] The radio map generation module 209 calls a neural network model to predict the radio signal strength of a specific area and obtain a radio map.

[0124] The specific implementation of each module in this embodiment can be found in Embodiment 1 above, and will not be repeated here. It should be noted that the device provided in this embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above.

[0125] Example 4

[0126] This embodiment provides a computer device, which can be a computer, such as... Figure 9 As shown, the system is connected via a system bus 901 to a processor 902, a memory, an input device 903, a display 904, and a network interface 905. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium 906 and internal memory 907. The non-volatile storage medium 906 stores the operating system, computer programs, and a database. The internal memory 907 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When the processor 902 executes the computer programs stored in the memory, it implements the federated learning-based radio map construction method proposed in Embodiment 1 above, as follows:

[0127] S1. Based on the characteristics of radio maps, design a neural network model for predicting radio signal strength;

[0128] S2. Obtain parameters including the number of user scheduling and the number of training rounds, and initialize the neural network model;

[0129] S3. Randomly schedule users according to the number of users scheduled in each round, and select users to participate in the federated learning in this round.

[0130] S4. The server compresses the neural network model and sends the compressed neural network model to the selected user.

[0131] S5. The selected user trains the model locally based on the model parameters sent by the server;

[0132] S6. The selected user compresses the parameter changes of the locally trained model and uploads the compressed parameter changes to the server.

[0133] S7. The server aggregates the parameter changes uploaded by the selected users and updates the model parameters based on the aggregated parameter changes.

[0134] S8. Based on the number of training rounds, repeat the steps of random user scheduling, downlink compressed transmission, user local training, uplink compressed transmission, and model aggregation update to complete the training of the neural network model.

[0135] S9. Call the neural network model to predict the radio signal strength of a specific area and obtain a radio map.

[0136] Example 5

[0137] This embodiment provides a storage medium, which is a computer-readable storage medium, storing a computer program. When the computer program is executed by a processor, it implements a federated learning-based radio map construction method of Embodiment 1 above, as follows:

[0138] S1. Based on the characteristics of radio maps, design a neural network model for predicting radio signal strength;

[0139] S2. Obtain parameters including the number of user scheduling and the number of training rounds, and initialize the neural network model;

[0140] S3. Randomly schedule users according to the number of users scheduled in each round, and select users to participate in the federated learning in this round.

[0141] S4. The server compresses the neural network model and sends the compressed neural network model to the selected user.

[0142] S5. The selected user trains the model locally based on the model parameters sent by the server;

[0143] S6. The selected user compresses the parameter changes of the locally trained model and uploads the compressed parameter changes to the server.

[0144] S7. The server aggregates the parameter changes uploaded by the selected users and updates the model parameters based on the aggregated parameter changes.

[0145] S8. Based on the number of training rounds, repeat the steps of random user scheduling, downlink compressed transmission, user local training, uplink compressed transmission, and model aggregation update to complete the training of the neural network model.

[0146] S9. Call the neural network model to predict the radio signal strength of a specific area and obtain a radio map.

[0147] The storage medium described in this embodiment can be a disk, optical disk, computer memory, random access memory (RAM), USB flash drive, portable hard drive, etc.

[0148] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for constructing radio maps based on federated learning, characterized in that, The radio map construction method includes: Based on the characteristics of radio maps, a neural network model for predicting radio signal strength was designed. Obtain parameters including the number of user scheduling and the number of training rounds, and initialize the neural network model; Users are randomly assigned based on the number of users scheduled in each round, and users are selected to participate in the federated learning in this round. The server compresses the neural network model and sends the compressed neural network model to the selected user. The selected users train the model locally based on the model parameters sent by the server; The selected user compresses the parameter changes of the locally trained model and uploads the compressed parameter changes to the server. The server aggregates the parameter changes uploaded by the selected users and updates the model parameters based on the aggregated parameter changes. Based on the number of training rounds, the steps of randomly scheduling users, downlink compressed transmission, user local training, uplink compressed transmission, and model aggregation and update are repeated to complete the training of the neural network model. A neural network model is used to predict the radio signal strength of a specific area, resulting in a radio map. The process by which the selected user compresses the parameter changes of the locally trained model and uploads the compressed parameter changes to the server is as follows: Calculate the change in model parameters based on the model parameters before and after local training. The number of unscheduled rounds is calculated based on the difference between the number of rounds the selected user had in the current round and the number of rounds the user was selected in the previous round. The uplink compression error feedback factor is then multiplied by the number of unscheduled rounds raised to the power of the error feedback coefficient. The vector to be compressed is obtained by adding the change in model parameters to the compression residual vector generated during the previous uplink compression and multiplying it by the error feedback coefficient. The number of uplink compressed quantization levels is determined based on the uplink compressed quantization bit depth. The vector to be compressed is multiplied by the compression scaling factor of the previous row and scaled. Elements whose absolute value after scaling is greater than the number of quantization levels minus one are limited to the number of quantization levels minus one. If the maximum absolute value after scaling is less than the number of quantization levels minus one, then multiply it by a coefficient so that its maximum absolute value is equal to the number of quantization levels minus one, and obtain the normalized vector. The normalized vector is then subjected to random rounding and denormalization in sequence to obtain the compressed model parameter change after the error feedback to be uploaded by the user, and this change is then uploaded to the server. Calculate the difference between the vector to be compressed and the model parameter change after the error feedback to be uploaded, obtain the compression residual vector generated by the row compression in this round, and record it locally on the user's device. The process by which the server aggregates the parameter changes uploaded by the selected users and updates the model parameters based on the aggregated parameter changes is as follows: The server averages the changes in model parameters uploaded by users in this round to obtain the aggregate vector for this round. The model update amount for this round is obtained by adding the aggregation vector of this round to the model update amount of the previous round multiplied by the momentum coefficient of the model aggregation. The server's model parameters are added to the model update amount in this round to obtain the aggregated model parameters.

2. The radio map construction method according to claim 1, characterized in that, The neural network model, based on the dimensions of the input data, the dimensions of the output predicted values, and the number of base stations involved in the radio map, includes a shared network and a personalized network. Among them, the shared network is used to predict the radio signal strength of all base stations; the number of personalized networks is consistent with the number of base stations involved in the radio map, and they are only responsible for predicting the radio signal strength of their corresponding base stations. The input data passes through the shared network and the personalized network in sequence. The personalized network outputs the predicted value of the radio signal strength of the corresponding base station.

3. The radio map construction method according to claim 1, characterized in that, The process involves acquiring relevant parameters, including the number of user scheduling attempts and the number of training rounds, and initializing the neural network model. The number of users scheduled is the number of users selected to participate in federated learning each round. The number of training rounds is the total number of times that the user is repeatedly randomly scheduled, downlink compressed transmission is performed, user local training is conducted, uplink compressed transmission is performed, and model aggregation updates are performed. The relevant parameters also include learning rate, data batch size, number of local iterations, uplink compression quantization bits, uplink compression scaling factor, uplink compression error feedback factor, downlink compression quantization bits, and momentum coefficient of model aggregation. The initialization of the neural network model involves randomly assigning initial values ​​to the model parameters.

4. The radio map construction method according to claim 1, characterized in that, The process by which the server compresses the model and sends the compressed model to the selected user is as follows: The number of downlink compressed quantization levels is determined based on the downlink compressed quantization bit depth. Based on the number of quantization levels and the maximum and minimum values ​​in the server's model parameters to be sent, the server's model parameters to be sent are subjected to min-max normalization, so that the numerical range of the absolute values ​​of the model parameters to be sent is normalized to be greater than or equal to zero and less than or equal to the number of downlink compressed quantization levels minus one, thus obtaining the normalized model parameter vector. The normalized model parameter vector is randomly rounded to obtain the parameters to be sent. The server then sends the parameters to be sent to the selected users. The parameters that have already been issued are denormalized to obtain the quantized model parameters; The model parameters on the server are updated to the quantized model parameters.

5. The radio map construction method according to claim 1, characterized in that, The selected user trains the neural network model locally based on the learning rate, data batch size, and local iteration number issued by the server.

6. A radio map construction system based on the federated learning method according to any one of claims 1 to 5, characterized in that, The radio map building system includes: The neural model design module is used to design a neural network model for predicting radio signal strength based on the characteristics of radio maps. The model initialization module is used to obtain parameters including the number of user scheduling and the number of training rounds, and to initialize the neural network model. The random user scheduling module is used to randomly schedule users based on the number of users scheduled in each round, and select users to participate in the current round of federated learning. The downlink compression transmission module is used by the server to compress the neural network model and send the compressed neural network model to the selected user; The user local training module is used by selected users to train the model locally based on the model parameters sent by the server. The uplink compression transmission module is used by the selected user to compress the parameter changes of the locally trained model and upload the compressed parameter changes to the server. The model aggregation and update module is used by the server to aggregate the parameter changes uploaded by the selected users and update the model parameters based on the aggregated parameter changes. The neural model training module repeats the steps of random user scheduling, downlink compressed transmission, user local training, uplink compressed transmission, and model aggregation update according to the number of training rounds to complete the training of the neural network model. The radio map generation module calls a neural network model to predict the radio signal strength of a specific area and obtain a radio map.

7. A computer device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the radio map construction method according to any one of claims 1-5.

8. A storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the radio map construction method according to any one of claims 1-5.

Images