CN114513270B - Heterogeneous wireless network spectrum resource sensing method and system based on federal learning - Google Patents

Heterogeneous wireless network spectrum resource sensing method and system based on federal learning Download PDF

Info

Publication number
CN114513270B
CN114513270B CN202210225145.6A CN202210225145A CN114513270B CN 114513270 B CN114513270 B CN 114513270B CN 202210225145 A CN202210225145 A CN 202210225145A CN 114513270 B CN114513270 B CN 114513270B
Authority
CN
China
Prior art keywords
global model
base station
model
local
base stations
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202210225145.6A
Other languages
Chinese (zh)
Other versions
CN114513270A (en
Inventor
盛洁
吴澄
朱巧明
李庆洋
张瑾
汪一鸣
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Suzhou University
Original Assignee
Suzhou University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Suzhou University filed Critical Suzhou University
Priority to CN202210225145.6A priority Critical patent/CN114513270B/en
Publication of CN114513270A publication Critical patent/CN114513270A/en
Application granted granted Critical
Publication of CN114513270B publication Critical patent/CN114513270B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/382—Monitoring; Testing of propagation channels for resource allocation, admission control or handover
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Abstract

The invention discloses a heterogeneous wireless network spectrum resource sensing method and system based on federal learning, which comprises the following steps: s1, sending a global model to all base stations in a region to which the global model belongs; s2, selecting partial base stations, and performing random gradient descent training on a local model of the selected base stations by using the global model to obtain a final local model; s3, obtaining a parameter difference value of the final local model and the global model, and updating the global model by federal aggregation; s4, repeating the steps S1-S3 to carry out iterative training on the global model until convergence, and obtaining a final global model; and S5, predicting the current occupation situation of all available spectrum resources in the region by using the parameters of the final global model. The invention greatly reduces the communication overhead and effectively prevents the leakage and illegal tampering of data.

Description

Heterogeneous wireless network spectrum resource sensing method and system based on federal learning
Technical Field
The invention relates to the technical field of communication, in particular to a heterogeneous wireless network spectrum resource sensing method and system based on federal learning.
Background
With the development of 5G communication and the progress of high-speed railways, mobile terminals and Internet of things equipment in high-speed railway transportation will increase explosively. In order to ensure that the quality of service is guaranteed and the communication requirements of the train, the trackside equipment and the passengers are better met, further optimization of the high-speed rail wireless communication resource management strategy is required.
Currently, in the field of railway wireless communication, LTE still faces many challenges. Moreover, the existing high-speed railway spectrum sharing mechanism also has a plurality of problems to be solved. The traditional spectrum sharing mechanism is centralized, but this generates a large amount of communication overhead and brings many security risks, such as single point of failure and denial of service attack. In addition, in the railway wireless communication system, the base stations are erected around the track and are mainly distributed in a chain manner, and the base stations form a chain heterogeneous network. How to realize the integration of spectrum resources in the chained heterogeneous network is a key challenge. In addition, the security and privacy of data are receiving more and more attention, and it is also a difficult point to ensure efficient security while sharing spectrum.
In 1999, joseph Mitola first proposed the concept of cognitive radio, and on this basis, in 2006 Ryan w. The cognitive radio network is a wireless communication network capable of perceiving the surrounding network environment and carrying out analysis and decision according to the perceived content, and can be self-adapted to the external complex dynamic network environment through the cognitive cycle, so that the network performance of end-to-end communication is improved. In the field of high-speed rail wireless communication, GSM-R, LTE-R and 5G-R exist simultaneously for a long time, and the cognitive radio network is of great help to improve the expandability and compatibility of the system. The cognitive radio technology provided by Ashwin Amanna for the first time can be applied to the field of rail transit, and the high-speed rail cognitive radio gets more and more attention and development. Compared with the traditional cognitive radio network, the high-speed rail cognitive radio network has new change. In the spectrum sensing stage, not only the video resource occupation condition of an authorized user needs to be sensed, but also the external track and physical environment condition need to be sensed; in a frequency spectrum decision stage, the method needs to adapt to the parameter change condition in the high-speed rail field and meet a specific optimization target; in the spectrum switching state, the traditional switching method is replaced by channel switching, trans-regional switching and parameter self-adaptive switching; in the spectrum sharing stage, due to the particularity of the high-speed rail wireless communication field, the secondary users generally do not act independently, so that the calculation of the channel capacity is simplified. However, the high-speed rail cognitive radio still has security threat, and only in the spectrum sensing stage, the high-speed rail cognitive radio can have the functions of imitating a master user attack, a terminal-spectrum sensing data tampering attack, eavesdropping on a control channel, a learning attack, a C-SSDF attack and the like.
The existing spectrum sharing technology is centralized, needs to be supported by a third-party organization, and generates massive communication overhead under the condition that a great number of users exist, which brings huge pressure to the network. In addition, there are various security issues with this approach, including single point of failure and denial of service attacks. The scene in the high-speed rail wireless communication system is special, the base stations are distributed in a chain manner and have isomerism, and how to sense and predict frequency spectrum resources in a chain heterogeneous network is a great difficulty. Moreover, the existing high-speed rail cognitive radio technology still needs to be continuously improved, the privacy security problem cannot be thoroughly solved, and the risks of data leakage and illegal tampering can be suffered.
Disclosure of Invention
The invention aims to provide a heterogeneous wireless network spectrum resource sensing method and system based on federal learning, which can greatly reduce communication overhead and effectively prevent data leakage and illegal tampering.
In order to solve the technical problem, the invention provides a heterogeneous wireless network spectrum resource sensing method based on federal learning, which comprises the following steps:
s1, sending a global model to all base stations in a region to which the global model belongs; the method specifically comprises the following steps:
s11, J base stations are arranged in the region where the global model belongs, each base station has a local data set, N samples are arranged in each local data set, and each sample consists of (N, x) 1 ，x 2 ，x 3 ，x 4 ，x 5 Y) composition;
where n represents the ID of a sample, x represents the characteristics of the sample, and x 1 Indicates the area, x, to which the base station belongs 2 Number, x, indicating the area to which the base station belongs 3 Indicates the type of the base station, x 4 Representing the number of the channel in the base station, selecting a time point as a time 0 point, x 5 Then represents the time t, y represents the label of the sample, and when the value of y is 0, it represents that the channel is in the idle state at the time; when the value of y is 1, the channel is in an occupied state at the moment;
s12, when the e-th global model is iterated, the global model G is processed e ＝(w e b e ) Sending to all base stations in the region, wherein w e ＝(w 1 ，w 2 ，w 3 ，w 4 ，w 5 ) Representing the weight taken up by each feature of the sample, b e Is a scalar, e represents the number of iterations performed by the global model;
s2, selecting partial base stations, and performing random gradient descent training on a local model of the selected base stations by using the global model to obtain a final local model; the method specifically comprises the following steps:
s21, randomly selecting M base stations from the I base stations in the area;
s22, the selected base station m receives the global model G e ＝(w e b e ) For local model
Figure GDA0003897687060000031
Performing local training, wherein the local training adopts a random gradient descent method to perform multiple iterative training, M is less than or equal to M, Q is less than or equal to Q, and Q represents the iteration times;
each local training process comprises the following steps:
s221, the base station m performs local model
Figure GDA0003897687060000032
Generating a preliminary prediction value of y for each sample thereof
Figure GDA0003897687060000033
Figure GDA0003897687060000034
S222, using sigmoid function to perform
Figure GDA0003897687060000035
Is compressed to [0,1 ]]In between, the expression is:
Figure GDA0003897687060000036
wherein p is the probability that the predicted value of y is 1, 0 represents idle, and 1 represents occupied;
the final predicted value
Figure GDA0003897687060000041
Figure GDA0003897687060000042
S223, the loss function of the base station m is:
Figure GDA0003897687060000043
wherein, y n Representing the true value, p, in the sample n n Representing predicted values for sample n
Figure GDA0003897687060000044
Probability of 1, N m Representing the total number of samples in the sample set;
s224, the local gradient value of the q-th iteration of the base station m is:
Figure GDA0003897687060000045
wherein the content of the first and second substances,
Figure GDA0003897687060000046
gradient operation is adopted, after the gradient is obtained, the local model is obtained by updating
Figure GDA0003897687060000047
S23, obtaining a final local model of the base station m after iteration is finished
Figure GDA0003897687060000048
S3, obtaining a parameter difference value of the final local model and the global model, and updating the global model by federal aggregation; the method specifically comprises the following steps:
obtaining the parameter difference between the local model and the global model
Figure GDA0003897687060000049
Performing a global aggregation operation, updating the global model according to the following formula:
Figure GDA00038976870600000410
wherein λ represents a learning rate;
s4, repeating the steps S1-S3 to carry out iterative training on the global model until convergence, and obtaining a final global model; the loss function of the global model for iterative training is as follows:
Figure GDA0003897687060000051
the convergence goal is to minimize the loss function:
O(w，b)＝arg min F (8)
obtaining parameters (w, b) of the final global model;
and S5, repeatedly utilizing the formulas (1), (2) and (3) by utilizing the parameters (w, b) of the final global model to predict the current occupation situation of all available spectrum resources in the region.
A heterogeneous wireless network spectrum resource sensing system based on federal learning comprises a central base station, wherein a plurality of base stations are arranged in the area of the central base station, and the central base station and the base stations form a communication link to jointly provide service for users; the central base station carries out spectrum resource sensing by adopting the method for sensing the spectrum resources of the heterogeneous wireless network based on the federal learning.
As a further development of the invention, the base stations are provided with edge mobile computing servers and the central base station is provided with a federal learning server.
As a further refinement of the present invention, the users include primary users and secondary users; the main user is an authorized user of the spectrum resource, and the owned idle spectrum is transacted with the base station of the nearby affiliated area; and the secondary user is an unauthorized user of the frequency spectrum resource and seeks network access service.
The invention has the beneficial effects that: the invention takes the heterogeneous base stations along the railway as the users for federal study, and each central base station as a server: the user database stores data about channel information, and a user only needs to upload trained parameter information to a server without uploading data, so that the communication overhead is greatly reduced, and data leakage and illegal tampering are effectively prevented; the server can predict the occupation conditions of the channels of all base stations in the system where the server is located according to the global model obtained after aggregation, so that spectrum resources are reasonably distributed, and users can share the spectrum while protecting the data privacy of the users; in addition, the federal learning needs a plurality of iterations to be converged, if all the participating users in each round send the training gradient information to the server, the communication overhead is increased, the problem is solved by the user scheduling method, namely, only part of the users are selected in each iteration to carry out local training, and the communication overhead is reduced.
Drawings
FIG. 1 is a schematic representation of a federated learning process according to an embodiment of the present invention;
FIG. 2 is a diagram of a system model according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a third simulation experiment scenario according to an embodiment of the present invention;
FIG. 4 is a diagram of the results of a third local training session according to an embodiment of the present invention;
FIG. 5 is a diagram of a triple Bandwidth learning accuracy result in accordance with an embodiment of the present invention;
FIG. 6 is a graph of the triple Bangbang learning loss results of the embodiment of the present invention.
Detailed Description
The present invention is further described below in conjunction with the following figures and specific examples so that those skilled in the art may better understand the present invention and practice it, but the examples are not intended to limit the present invention.
The invention provides a heterogeneous wireless network spectrum resource sensing method based on federal learning, which comprises the following steps:
s1, sending the global model to all base stations in the area to which the global model belongs;
s2, selecting partial base stations, and performing random gradient descent training on a local model of the selected base stations by using the global model to obtain a final local model;
s3, obtaining a parameter difference value of the final local model and the global model, and updating the global model by federal aggregation;
s4, repeating the steps S1-S3 to carry out iterative training on the global model until convergence, and obtaining a final global model;
and S5, predicting the current occupation situation of all available spectrum resources in the region by using the parameters of the final global model.
Federal learning is a promising approach to edge intelligence to protect privacy in distributed scenarios. In traditional machine learning, the server would collect all training data, and federal learning largely solved this problem. The user would first download the latest global model information and then train the data locally. In traditional federal learning, a user can train gradient information according to own data during local training, and the data is not required to be uploaded to a server, but only the gradient information is sent to the server. The server will aggregate the gradient information of multiple users to update the global model parameters and send the latest global model to the participating users. The invention takes the heterogeneous base stations along the railway as users and each central base station as a server. The user database stores data related to channel information, and a user only needs to upload trained parameter information to the server without uploading the data, so that the communication overhead is greatly reduced, and the data leakage and illegal tampering are effectively prevented. The server can predict the occupation status of the channels of all base stations in the system where the server is located according to the global model obtained after aggregation, so that spectrum resources are reasonably distributed, and users can share the spectrum while protecting the data privacy of the users. The performance of federal learning is limited due to limited available wireless spectrum resources and insufficient computing power of the users themselves. In addition, federal learning requires multiple iterations to converge, which increases communication overhead if all participating users in each round send their own training gradient information to the server. The invention solves the problem by a user scheduling method, namely only part of users are selected for local training in each iteration, which reduces the communication overhead.
Example one
Based on the above embodiment, assuming that there are J base stations in the area i and numbering the base stations, the base stations J may use a set
Figure GDA0003897687060000071
And (4) showing. For the
Figure GDA0003897687060000072
All have a local data set
Figure GDA0003897687060000073
Each one of which
Figure GDA0003897687060000074
In which is N j Samples, each sample representing the occupancy state of a channel of the base station at a time in the past, are represented by (n, x) 1 ，x 2 ，x 3 ，x 4 ，x 5 Y) where n represents the ID of the sample, x represents the characteristics of the sample, and x 1 Indicates the area to which the base station belongs, x 2 Number, x, indicating the area to which the base station belongs 3 Indicates the type of the base station, x 4 Representing the number of the channel in the base station, and selecting a time point as a time point 0, x 5 Representing the time t. y represents the label of the sample, and when the value of y is 0, the channel is in an idle state at the moment; a value of y of 1 indicates that the channel is occupied at that time.
Such as
Figure GDA0003897687060000075
It is the data set for base station # 3 in area 2, which is a class II baseAnd (4) a station. It has a total of 160 samples from 0 to 159. Sample 0 indicates that channel 0 is in the occupied state at time 4 and sample 159 indicates that channel 9 is in the idle state at time 8.
Local model training and global model aggregation for federal learning:
in a coverage area of a certain base station group, sensing and predicting of frequency spectrum resources are achieved by using a federal learning algorithm. In the invention, the federated learning algorithm comprises a global model broadcast, a training node selection, a local model training and a global model aggregation in each iteration process.
Suppose that at the e-th algorithm iteration, the server base station (central base station) in the area first maps the global model G e To users in the area (including three types of common base stations, i.e., client base stations), where G e ＝(w e b e )，w e ＝(w 1 ，w 2 ，w 3 ，w 4 ，w 5 ) Representing the weight of each feature of the sample, b e Is a scalar. Then, the server base station randomly selects M base stations from the J base stations in the area to serve as a training set of the spectrum resource perception model based on the federal learning, and the remaining (J-M) base stations serve as a test set of the model to be used for verifying the performance of the model. On the basis, the federal learning local model training and the global model aggregation are developed.
Local model training: when the selected base station receives the global model, local training is started, and the process adopts a random gradient descent method for training, which is also a process of multiple iterations. Each local iteration process q will comprise the following process.
User m utilizes the updated local model
Figure GDA0003897687060000081
Training data of which
Figure GDA0003897687060000082
When q =0, the ratio of q to q is set to zero,
Figure GDA0003897687060000083
user m will generate a preliminary prediction of y for each of its samples according to the local model
Figure GDA0003897687060000084
Figure GDA0003897687060000084
0 indicates idle and 1 indicates occupied. The specific process is as follows:
Figure GDA0003897687060000085
wherein, the first and the second end of the pipe are connected with each other,
Figure GDA0003897687060000086
the value range is not in [0,1 ]]In the method, sigmoid function is required to be used
Figure GDA0003897687060000087
Is compressed to [0,1 ]]In the meantime. The specific expression is as follows:
Figure GDA0003897687060000088
where p can be understood as the probability that the predicted value of y is 1, the final predicted value
Figure GDA0003897687060000089
Figure GDA00038976870600000810
The loss function for user m can be expressed as:
Figure GDA0003897687060000091
wherein, y n Representing the true value, p, in the sample n n Representing predicted values for sample n
Figure GDA0003897687060000092
A probability of 1.
The local gradient values for user m are:
Figure GDA0003897687060000093
wherein the content of the first and second substances,
Figure GDA0003897687060000094
is a gradient operation. After obtaining the gradient, the local model is updated
Figure GDA0003897687060000095
After the iteration is completed, the final local model can be obtained
Figure GDA0003897687060000096
Will be different value
Figure GDA0003897687060000097
And sending the data to a server.
And (3) global model aggregation: after receiving the m pieces of parameter information, the server executes global aggregation operation, and updates the global model according to a formula (6):
Figure GDA0003897687060000098
where λ represents a learning rate.
In the present invention, the global loss function in the global iterative process e can be represented as:
Figure GDA0003897687060000099
the invention improves the training accuracy rate by the loss function reduction, and the aim is to minimize the loss function, namely:
O(w，b)＝arg min F (8)
the spectrum resource sensing algorithm of the chain heterogeneous wireless network based on the federal learning comprises the following steps:if a sub-user s is in a certain area req To a nearby central base station c i Application for access service, c i It is necessary to know the spectrum information of the base stations in this area and even other areas for s req The central base station is required to interact with a general base station in the area in real time, and the current occupation situation of all available spectrum resources in the area is sensed and predicted by utilizing the prior spectrum occupation situation. And sensing and predicting the frequency spectrum in the region by adopting a federal learning algorithm. The parameters (w, b) can be obtained based on the formula (8), and then the current occupation situation of all available spectrum resources in the region can be predicted by repeatedly using the formulas (1), (2) and (3). The whole spectrum resource sensing algorithm of the chain heterogeneous wireless network is shown in table 1, and the flow chart is shown in fig. 1.
Table 1:
Figure GDA0003897687060000101
example two
The embodiment provides a heterogeneous wireless network spectrum resource sensing system based on federal learning, which divides a certain section of railway link type wireless communication network into I areas, and the areas cover the section of network. Communication bodies of 4 types of parameters and spectrum resource management are defined in each area, and comprise Primary Users (PU), secondary Users (SU), base Stations (BS) and Central Base Stations (CBS). With S = { S = i SU, which refers to an unauthorized user of spectrum resources, i.e., communication devices on the train, and mobile terminals of passengers, seeking network access services. With P = { P i Define PU, which is an authorized user of spectrum resources, and to obtain revenue, PU will trade the free spectrum owned in hand with nearby BSs. By using
Figure GDA0003897687060000102
To define the jth BS of the ith zone. Each BS is equipped with an edge mobile computing (MEC) server with certain computing and communication capabilities, forming an uplink with the CBS and a downlink with the PU and SU. With C = { C i Define CBS in the entire link. Each CBS is equipped with a federal learning server (FL) and has extremely strong computing and communication capabilities. The schematic diagram of the system model is shown in fig. 2, in which the train represents SU and the large base station at the center of the lower row represents CBS in a certain area. The other small grey base stations (with and without background grey boxes) represent three different types of common BSs, which are distributed in chains along the railway. After the PU and the BS complete the transaction, the BS manages the spectrum resources of the transaction, so the PU is not shown in the figure, but the PU is also an important component of the whole system.
The system adopts the spectrum resource sensing method of the heterogeneous wireless network based on the federal learning as the first implementation mode or embodiment to sense the spectrum resource.
EXAMPLE III
In the embodiment, a chain heterogeneous network environment is built on a python3.8 software platform to carry out simulation experiments. An experimental scenario was designed as shown in fig. 3. In the figure, one large circle indicates an area in which a straight solid line indicates a railroad link, o indicates a class I base station, x indicates a class ii base station, a-indicates a class iii base station, and ^ indicates a central server. The data sets are stored in the form of tuples in the database of each base station. The detailed parameters of the three types of base stations are shown in table 2.
Table 2:
Figure GDA0003897687060000111
local model convergence and spectrum sensing accuracy experiment: as shown in fig. 4, the result is obtained by local training of a class I base station in a global iteration, and it can be seen from the figure that when the number of local iterations increases, the accuracy of predicting channel occupancy continuously increases and rapidly converges, and gradually stabilizes after reaching 0.9 or more, and the cross entropy loss thereof also continuously decreases and tends to stabilize after reaching 0.41, thereby achieving convergence.
The federal learning global aggregation model convergence and perception accuracy experiment comprises: the federally learned simulation parameters for a region are shown in table 3. When one part of base stations are selected in the area for federal learning, the other part of the rest base stations are used as a test set for verifying the accuracy of federal learning prediction.
Table 3:
number of areas 1
Number of repeated tests 10
Number of global iterations 40
Number of class I base stations participating in federal learning 12
Number of class II base stations participating in federal learning 12
Number of class III base stations participating in federal learning 4
Number of base stations selected per global iteration 3/6/12
The simulation results of the federal learning are shown in fig. 5 and fig. 6, in which "m =3", "m =6" and "m =12" in the legend represent the number of users selected at each global iteration of the federal learning.
Fig. 5 shows the average accuracy of the simulation results when the number of nodes participating in the training is different after a plurality of tests, and it is obvious from the figure that the larger the number of nodes is, the faster the convergence speed of the accuracy is, and the fewer the training times required for achieving the convergence are. However, the three converged are almost the same in accuracy, and eventually approach 0.98. The results were the same as expected. Note also that the more nodes that participate in the local training, the more communication overhead in the network will increase, but correspondingly, the less computation overhead will be required by the central base station.
Fig. 6 shows the average loss of simulation results when the number of nodes participating in training is different after multiple experiments. It is obvious from the figure that the larger the number of the participating nodes is, the faster the convergence speed is lost, and the less the training times are required for reaching the convergence. However, the three converged accuracies are almost the same, and eventually approach 0.1.
The above-mentioned embodiments are merely preferred embodiments for fully illustrating the present invention, and the scope of the present invention is not limited thereto. The equivalent substitution or change made by the technical personnel in the technical field on the basis of the invention is all within the protection scope of the invention. The protection scope of the invention is subject to the claims.

Claims (4)

1. A heterogeneous wireless network spectrum resource sensing method based on federal learning is characterized in that: the method comprises the following steps:
s1, sending the global model to all base stations in the area to which the global model belongs; the method specifically comprises the following steps:
s11, J base stations are arranged in the region where the global model belongs, each base station has a local data set, N samples are arranged in each local data set, and each sample is composed of (N, x) 1 ，x 2 ，x 3 ，x 4 ，x 5 Y) composition;
where n represents the ID of a sample, x represents the characteristics of the sample, and x 1 Indicates the area, x, to which the base station belongs 2 Number, x, indicating the area to which the base station belongs 3 Indicates the type of the base station, x 4 Representing the number of the channel in the base station, and selecting a time point as a time point 0, x 5 Then it represents the time of dayt, y represents the label of the sample, and when the value of y is 0, the channel is in an idle state at the moment; when the value of y is 1, the channel is in an occupied state at the moment;
s12, when the e-th global model is iterated, the global model G is used e ＝(W e b e ) Sending to all base stations in the region, wherein w e ＝(w 1 ，w 2 ，w 3 ，w 4 ，w 5 ) Representing the weight of each feature of the sample, b e Is a scalar, e represents the number of iterations performed by the global model;
s2, selecting partial base stations, and performing random gradient descent training on a local model of the selected base stations by using the global model to obtain a final local model; the method specifically comprises the following steps:
s21, randomly selecting M base stations from the J base stations in the area;
s22, the selected base station m receives the global model G e ＝(w e b e ) For local model
Figure FDA0003897687050000011
Performing local training, wherein the local training adopts a random gradient descent method to perform multiple iterative training, M is less than or equal to M, Q is less than or equal to Q, and Q represents the number of iterations;
each local training process comprises the following steps:
s221, the base station m performs local model
Figure FDA0003897687050000012
Generating a preliminary prediction of y for each of its samples
Figure FDA0003897687050000021
Figure FDA0003897687050000022
S222, using sigmoid function to convert
Figure FDA0003897687050000023
Is compressed to [0,1 ]]In between, the expression is:
Figure FDA0003897687050000024
wherein p is the probability that the predicted value of y is 1, 0 represents idle, and 1 represents occupied;
the final predicted value
Figure FDA0003897687050000025
Figure FDA0003897687050000026
S223, the loss function of the base station m is as follows:
Figure FDA0003897687050000027
wherein, y n Representing the true value, p, in the sample n n Representing predicted values for sample n
Figure FDA0003897687050000028
Probability of 1, N m Representing the total number of samples in the sample set;
s224, the local gradient value of the q-th iteration of the base station m is:
Figure FDA0003897687050000029
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA00038976870500000210
is gradient operation, after obtaining the gradient, the update is obtainedLocal model
Figure FDA00038976870500000211
S23, obtaining a final local model of the base station m after iteration is finished
Figure FDA00038976870500000212
S3, obtaining a parameter difference value of the final local model and the global model, and updating the global model by federal aggregation; the method specifically comprises the following steps:
obtaining the parameter difference value of the local model and the global model
Figure FDA0003897687050000031
Performing a global aggregation operation, updating the global model according to the following formula:
Figure FDA0003897687050000032
wherein λ represents a learning rate;
s4, repeating the steps S1-S3 to carry out iterative training on the global model until convergence, and obtaining a final global model; the loss function of the global model for iterative training is as follows:
Figure FDA0003897687050000033
the convergence goal is to minimize the loss function:
O(w，b)＝arg min F (8)
obtaining parameters (w, b) of a final global model;
and S5, repeatedly utilizing the formulas (1), (2) and (3) by utilizing the parameters (w, b) of the final global model to predict the current occupation situation of all available spectrum resources in the region.
2. A heterogeneous wireless network spectrum resource sensing system based on federal learning is characterized in that: the system comprises a central base station, wherein a plurality of base stations are arranged in the area of the central base station, and the central base station and the base stations form a communication link to jointly provide service for users; the central base station performs spectrum resource sensing by using the spectrum resource sensing method of the heterogeneous wireless network based on federal learning according to claim 1.
3. The system for sensing spectrum resources of a heterogeneous wireless network based on federated learning of claim 2, wherein: the base station is provided with an edge mobile computing server, and the central base station is provided with a federal learning server.
4. The system for sensing spectrum resources of a heterogeneous wireless network based on federated learning of claim 2, wherein: the users comprise primary users and secondary users; the main user is an authorized user of the spectrum resource, and the owned idle spectrum is transacted with the base station of the nearby affiliated area; and the secondary user is an unauthorized user of the frequency spectrum resource and seeks network access service.
CN202210225145.6A 2022-03-07 2022-03-07 Heterogeneous wireless network spectrum resource sensing method and system based on federal learning Active CN114513270B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210225145.6A CN114513270B (en) 2022-03-07 2022-03-07 Heterogeneous wireless network spectrum resource sensing method and system based on federal learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210225145.6A CN114513270B (en) 2022-03-07 2022-03-07 Heterogeneous wireless network spectrum resource sensing method and system based on federal learning

Publications (2)

Publication Number Publication Date
CN114513270A CN114513270A (en) 2022-05-17
CN114513270B true CN114513270B (en) 2022-12-02

Family

ID=81554384

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210225145.6A Active CN114513270B (en) 2022-03-07 2022-03-07 Heterogeneous wireless network spectrum resource sensing method and system based on federal learning

Country Status (1)

Country Link
CN (1) CN114513270B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117676594A (en) * 2022-08-10 2024-03-08 索尼集团公司 Electronic device and method for spectrum management apparatus, and storage medium
CN116318465B (en) * 2023-05-25 2023-08-29 广州南方卫星导航仪器有限公司 Edge computing method and system in multi-source heterogeneous network environment

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113011599A (en) * 2021-03-23 2021-06-22 上海嗨普智能信息科技股份有限公司 Federal learning system based on heterogeneous data
CN113038616A (en) * 2021-03-16 2021-06-25 电子科技大学 Frequency spectrum resource management and allocation method based on federal learning
CN114125861A (en) * 2021-11-26 2022-03-01 北京邮电大学 Wireless federal learning method and device
WO2022041947A1 (en) * 2020-08-24 2022-03-03 华为技术有限公司 Method for updating machine learning model, and communication apparatus

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111754000B (en) * 2020-06-24 2022-10-14 清华大学 Quality-aware edge intelligent federal learning method and system

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2022041947A1 (en) * 2020-08-24 2022-03-03 华为技术有限公司 Method for updating machine learning model, and communication apparatus
CN113038616A (en) * 2021-03-16 2021-06-25 电子科技大学 Frequency spectrum resource management and allocation method based on federal learning
CN113011599A (en) * 2021-03-23 2021-06-22 上海嗨普智能信息科技股份有限公司 Federal learning system based on heterogeneous data
CN114125861A (en) * 2021-11-26 2022-03-01 北京邮电大学 Wireless federal learning method and device

Also Published As

Publication number Publication date
CN114513270A (en) 2022-05-17

Similar Documents

Publication Publication Date Title
Seid et al. Collaborative computation offloading and resource allocation in multi-UAV-assisted IoT networks: A deep reinforcement learning approach
CN114513270B (en) Heterogeneous wireless network spectrum resource sensing method and system based on federal learning
Hu et al. A deep reinforcement learning-based framework for dynamic resource allocation in multibeam satellite systems
Sun et al. Radio resource management for D2D-based V2V communication
Li et al. A dynamic graph optimization framework for multihop device-to-device communication underlaying cellular networks
Lv et al. Beyond 5G for digital twins of UAVs
CN110691406A (en) D2D frequency spectrum efficient sharing method for internet-connected unmanned safe communication
CN107659915A (en) Car networking people's car resource allocation methods based on geographical area information
Mensah et al. Energy efficiency based joint cell selection and power allocation scheme for HetNets
Hao et al. URLLC resource slicing and scheduling for trustworthy 6G vehicular services: A federated reinforcement learning approach
Mandloi et al. Seamless connectivity with 5G enabled unmanned aerial vehicles base stations using machine programming approach
Ho et al. Optimized base station allocation for platooning vehicles underway by using deep learning algorithm based on 5G-V2X
Wang et al. Covert communications in air-ground integrated urban sensing networks enhanced by federated learning
Ni et al. An effective hybrid V2V/V2I transmission latency method based on LSTM neural network
Hua et al. On sum-rate maximization in downlink UAV-aided RSMA systems
Qi et al. Social prediction-based handover in collaborative-edge-computing-enabled vehicular networks
Huang et al. Performance of resource allocation in device-to-device communication systems based on particle swarm optimization
Hossen et al. Relax online resource allocation algorithms for D2D communication
Yan et al. Joint resource allocation and power control for V2V communication of high-density vehicle network
CN108551676A (en) A kind of cognition car networking frequency spectrum distributing method based on immune optimization
YANG et al. Enabling Intelligence at Network Edge: An Overview of Federated Learning
Wang et al. Deep reinforcement learning-based optimization for end-to-end network slicing with control-and user-plane separation
Bechihi et al. A resource allocation algorithm for formation control of connected vehicles
CN103812629B (en) A kind of resource allocation methods under high ferro cell on wheels communication construction
Mardan et al. Using AI in wireless communication system for resource management and optimisation

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant