CN112911608A - Large-scale access method for edge-oriented intelligent network - Google Patents

Large-scale access method for edge-oriented intelligent network Download PDF

Info

Publication number
CN112911608A
CN112911608A CN202110049384.6A CN202110049384A CN112911608A CN 112911608 A CN112911608 A CN 112911608A CN 202110049384 A CN202110049384 A CN 202110049384A CN 112911608 A CN112911608 A CN 112911608A
Authority
CN
China
Prior art keywords
base station
model
global model
mobile
mobile device
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.)
Granted
Application number
CN202110049384.6A
Other languages
Chinese (zh)
Other versions
CN112911608B (en
Inventor
齐俏
陈晓明
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.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
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 Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202110049384.6A priority Critical patent/CN112911608B/en
Publication of CN112911608A publication Critical patent/CN112911608A/en
Application granted granted Critical
Publication of CN112911608B publication Critical patent/CN112911608B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W16/00—Network planning, e.g. coverage or traffic planning tools; Network deployment, e.g. resource partitioning or cells structures
    • H04W16/22—Traffic simulation tools or models
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L25/00—Baseband systems
    • H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
    • H04L25/0202—Channel estimation
    • H04L25/0204—Channel estimation of multiple channels
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W16/00—Network planning, e.g. coverage or traffic planning tools; Network deployment, e.g. resource partitioning or cells structures
    • H04W16/24—Cell structures
    • H04W16/28—Cell structures using beam steering
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/06—Testing, supervising or monitoring using simulated traffic

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Theoretical Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Power Engineering (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

The invention discloses a large-scale access method facing an edge intelligent network. A cell deploys a multi-antenna base station as an edge server, and a large number of mobile devices access a wireless network through the base station for federal learning. In each round of learning, the base station firstly selects partial equipment and broadcasts a global model through a downlink design transmission beam. And after each selected device recovers the global model through the receiver, a new local model is trained based on the local data set, and then the new local model is transmitted after beam forming. By using the superposition characteristics of the wireless channel, the base station can directly obtain the aggregation model calculated in the air through a receiver, and then calculate the weighted average of the aggregation model as an updated global model. The invention provides an effective large-scale access method for the edge intelligent network.

Description

Large-scale access method for edge-oriented intelligent network
Technical Field
The invention relates to the field of wireless communication, in particular to a large-scale access method facing to an edge intelligent network.
Background
In recent years, with the advent of the world of everything interconnection, the number of mobile devices and the amount of data generated by them has increased explosively. Cisco release reports indicate that the number of globally networked devices in 2021 will increase from 171 billion to 271 billion in 2016, and the amount of data generated by the devices will also increase from 218ZB to 847ZB in 2016. In the face of such huge equipment quantity and massive data, the Internet of things can combine advanced information processing technologies such as data mining, artificial intelligence and the like to extract useful information for calculation, analysis and processing, so that information interaction and seamless connection between people and objects and between objects are realized, and various intelligent Internet of things services and applications are enabled.
However, the realization of the 'everything intelligent connection' assisted by the artificial intelligence technology in the 6G era still faces some problems and challenges. Such as transmitting data with private information to a cloud server, increases the risk of revealing user private data. In addition, due to limited spectrum resources, model learning by aggregating massive distributed data over a wireless channel can cause network congestion and network delay. Based on this, edge intelligence arises, and because data acquisition and processing are mainly performed at the network edge, edge intelligence can significantly reduce service delay and energy consumption of edge equipment, slow down the pressure of network bandwidth, and improve the processing efficiency of data in the era of intellectual association.
Federal learning is a promising solution in edge intelligence, can effectively address the need for privacy-sensitive and low-latency internet of things, and has the ability to utilize distributed computing resources. However, when the number of edge devices is large, the large communication delay becomes a major factor limiting the edge intelligence. To solve the above problems, the advent of over-the-air computation has made it possible for large-scale edge devices to learn with high spectral efficiency and low latency, given the limited communication bandwidth and stringent latency requirements. Specifically, federal learning allows each mobile device to store its data locally, and when model training is performed through a wireless link, each mobile device only needs to upload its locally updated model to the edge server aggregation, which can effectively prevent the data collected by the mobile device from being leaked to other devices and the edge server, thereby enhancing the privacy and data security of the device. By coordinating the local models of the mobile devices participating in learning, the edge server can directly aggregate through over-the-air computation to obtain a high-precision shared global model.
Therefore, the combination of the over-the-air calculation and the federal learning technology is expected to break through the bottleneck problem of communication in the edge intelligent network and solve a series of problems of large-scale mobile equipment accessing the wireless network.
Disclosure of Invention
The invention provides a large-scale access method facing an edge intelligent network, aiming at solving the problems of limited communication, low calculation efficiency, overhigh time delay, low spectrum efficiency and the like when large-scale mobile equipment is accessed in the scheme.
The invention adopts the following specific technical scheme:
a large-scale access method facing to an edge intelligent network comprises the following steps:
s1: a base station with N antennas is deployed in advance, mobile equipment with K single antennas is accessed into a wireless network through the base station to perform federal learning, and a global model is regarded as one round of learning when the global model is updated every time; channel state information from the base station to the kth mobile device is obtained through channel estimation or feedback
Figure BDA0002898677310000021
The sequence number K is 1, …, K, and the real channel state information is
Figure BDA0002898677310000022
Wherein ekFor estimating the error vector for the channel, the norm of which has a boundary value epsilonkSatisfy | | | ek||≤εk(ii) a Initializing the number t of learning rounds to be 1;
s2: in the t-th round of learning, the base station performs ascending sequencing on all the mobile devices based on the channel state information to obtain a selected priority reordering xπ(1)≤…≤xπ(K)Wherein with xk∈[0,1]Representing weights of a kth mobile device, said selected priority reordering being according to xkIs ordered by the value size of xkThe smaller the value of (A) is, the higher the priority is selected, and the higher the ranking is; pi (1) is a serial number corresponding to the mobile equipment with the minimum weight value, pi (i) is a serial number corresponding to the ith mobile equipment which is ranked in the front after the weights of K mobile equipment are ranked from small to large, and pi (K) is a serial number corresponding to the mobile equipment with the maximum weight value; the values of pi (1), … and pi (K) are respectively one of 1, … and K and are not repeated;
s3: base station is reordered x according to selected priority of equipmentπ(1)≤…≤xπ(K)Performing feasibility detection to obtain the selected equipment set S in the t-th round of learningtPi (1), pi (2), …, pi (m) }, where m ∈ [1, K]Is the maximum value that makes the device set feasible, and then the transmit beam w, receive beam z of the base station and the receiver v of the selected mobile device k are designedkAnd a transmitter bk；
S4: base station through downlink via transmitting beam w to selected device set StBroadcasting the updated global model q[t -1]；
S5: the selected mobile device k belongs to StBy receivers vkRecovering a received global model
Figure BDA0002898677310000031
And training local data set DkObtaining an updated local model
Figure BDA0002898677310000032
Then via transmitter bkSimultaneously transmitting to the base station through the uplink;
s6: after receiving the signals transmitted by the mobile equipment, the base station recovers the aggregated local model through the receiving beam z, and calculates the weighted average of the local models of all the mobile equipment as an updated global model q[t]；
S7: judging a global model q[t]Whether the global model is converged or not, if the global model is converged, the global model q is output[t]Otherwise, let t be t +1, the execution is resumed from step S2.
Based on the technical scheme, part of the steps can be realized in the following preferred mode.
Preferably, in step S2, the method for the base station to rank all the mobile devices based on the channel state information to obtain the selected priority comprises:
s21: transmitting wave beam when initializing base station broadcasting global model
Figure BDA0002898677310000033
Receiver v of mobile device kkTransmitter when mobile device k uploads local model 1
Figure BDA0002898677310000034
Base station receiving beam z ═ 1,0, …,0]TIn which P isBSFor the maximum transmission power, P, of the base stationkIs the maximum transmit power of mobile device k;
s22: calculating the mean square error between the global model estimated by the device k and the global model broadcasted by the base station when the base station downlink broadcasts
Figure BDA0002898677310000035
Calculating the mean square error between the model of base station aggregation and the target summation model when the model is uploaded by the equipment
Figure BDA0002898677310000036
Wherein
Figure BDA0002898677310000037
The variance of Gaussian white noise, | - | represents the absolute value of a scalar, and | | - | | represents the norm of a vector;
s23: introducing an auxiliary variable eta according to the mean square error limit of the base station and the mobile equipmentk≥0,ρk≥0,φk≥0,
Figure BDA0002898677310000038
Order to
Figure BDA0002898677310000039
Figure BDA0002898677310000041
Where γ is the base station allowed maximum mean square error bound, δkIs the maximum mean square error limit allowed by the mobile device k, I is the identity matrix with dimension N;
s24: based on maximum transmit power limits of the base station and the mobile device, the method comprises
Figure BDA0002898677310000042
And | | w | | non-conducting phosphor2≤PBS；
S25: solve by alternating iterations
Figure BDA0002898677310000043
Get the weight value x ═ x of all mobile devices1,…,xK]T；
S26: determining target value | x | non-woven phosphor1Whether to converge or not, if the target value | | x | | non-woven phosphor1And when the convergence occurs, the mobile devices are reordered according to the priority selected by the current weight values x of all the mobile devices, otherwise, the step S22 is executed again.
Preferably, in step S3, the selected device set S is obtained by feasibility detectiontDesigning a base station transmitting beam w, a receiving beam z and a receiver v of a selected mobile device kkAnd a transmitter bkThe method comprises the following steps:
s31: the initialization parameter m is equal to K,
Figure BDA0002898677310000044
vk＝1，
Figure BDA0002898677310000045
z＝[1,0,…,0]T；
s32: reordering x according to device selected priorityπ(1)≤…≤xπ(K)The selected device set is St＝{π(1),π(2),…,π(m)}；
S33: when the downlink broadcast of the base station is calculated, the selected mobile equipment k belongs to StMean square error between estimated global model and global model broadcast by base station
Figure BDA0002898677310000046
Calculating the mean square error between the model aggregated by the base station and the target summation model when the selected mobile equipment uploads the local model
Figure BDA0002898677310000047
S34: according to the base station and the selected mobile equipment k belonging to StMean square error limitation of, introducing an auxiliary variable ηk≥0,ρk≥0,φk≥0,
Figure BDA0002898677310000048
Order to
Figure BDA0002898677310000049
Figure BDA00028986773100000410
S35: according to the base station and the selected mobile equipment k ∈ StMaximum transmit power limit of
Figure BDA0002898677310000051
And | | w | | non-conducting phosphor2≤PBS；
S36: solve by alternating iterations
Figure BDA0002898677310000052
The solution with the smallest value of (c); if the solution exists, detecting the value of the mean square error of the base station, and if the solution satisfies the condition
Figure BDA0002898677310000053
To obtain w, vk,bkAnd z, and outputting the value of m, the selected device is St-pi (1), pi (2), …, pi (m) }; if no solution or not satisfied
Figure BDA0002898677310000054
Let m be m-1 and the execution start again from step S32.
Preferably, in step S5, the method for updating the local model includes: training local data set D using convolutional neural networkkThe gradient updating method in the network is a random gradient descent method, and the local model updating formula is
Figure BDA0002898677310000055
Wherein
Figure BDA0002898677310000056
Is a gradient operation, Lk(. cndot.) is a loss function, and θ is the learning rate.
Preferably, in S25 and S36, the alternating iterative solution is performed by using an interior point method or directly calling a CVX toolkit.
The invention has the beneficial effects that: the edge-intelligent-oriented large-scale access method provided by the invention solves a series of problems caused by data privacy leakage, limited communication and overhigh calculation delay due to the fact that massive mobile equipment is accessed into a wireless network. The equipment selection method and the transceiver design method based on the channel state information design have the advantages of low calculation time delay, high spectrum efficiency, effective interference suppression and the like.
Drawings
FIG. 1 is a system block diagram of a large scale access method for an edge-oriented intelligent network;
fig. 2 is a comparison of the performance of the proposed method in case of different number of antennas at the base station and different channel estimation error bounds (number of antennas 32, 48 and 64, respectively);
fig. 3 shows the comparison of testing accuracy of federal learning under different access methods (i.e. the ideal case of no error of all-selected devices, the proposed robust access method and the non-robust access method).
Detailed Description
In this embodiment, a system block diagram of a large-scale access method for an edge intelligent network is shown in fig. 1, where a base station has N antennas, and K mobile devices with single antennas access a wireless network through the base station to perform federal learning. In each round of learning, the base station firstly selects partial equipment according to the channel state information, and broadcasts a global model through a downlink design transmission beam. And after each selected device recovers the global model through the receiver, a new local model is trained based on the local data set, and then the new local model is transmitted after beam forming. By using the superposition characteristics of the wireless channel, the base station can directly obtain the aggregation model calculated in the air through a receiver, and then calculate the weighted average of the aggregation model as an updated global model.
The specific technical scheme adopted by the embodiment is as follows:
the large-scale access method facing the edge intelligent network comprises the following steps:
s1: a base station with N antennas is deployed in advance, mobile equipment with K single antennas is accessed into a wireless network through the base station to perform federal learning, and a global model is regarded as one round of learning when the global model is updated every time; channel state information from the base station to the kth mobile device is obtained through channel estimation or feedback
Figure BDA0002898677310000061
The sequence number K is 1, …, K, and the real channel state information is
Figure BDA0002898677310000062
Wherein ekFor estimating the error vector for the channel, the norm of which has a boundary value epsilonkSatisfy | | | ek||≤εk(ii) a The number of initial learning rounds t is 1.
S2: in the t-th round of learning, the base station performs ascending sequencing on all the mobile devices based on the channel state information to obtain a selected priority reordering xπ(1)≤…≤xπ(K)Wherein with xk∈[0,1]Representing weights of a kth mobile device, said selected priority reordering being according to xkIs ordered by the value size of xkThe smaller the value of (A) is, the higher the priority is selected, and the higher the ranking is; pi (1) is the minimum weight valueThe serial number corresponding to the mobile device of (1), pi (i) is the serial number corresponding to the ith mobile device which is ranked earlier after the weights of the K mobile devices are ranked from small to large, and pi (K) is the serial number corresponding to the mobile device with the largest weight value; the values of pi (1), … and pi (K) are respectively one of 1, … and K and are not repeated.
In this step, the method for the base station to rank all the mobile devices based on the channel state information to obtain the selected priority ranking is as follows:
s21: transmitting wave beam when initializing base station broadcasting global model
Figure BDA0002898677310000063
Receiver v of mobile device kkTransmitter when mobile device k uploads local model 1
Figure BDA0002898677310000064
Base station receiving beam z ═ 1,0, …,0]TIn which P isBSFor the maximum transmission power, P, of the base stationkIs the maximum transmit power of mobile device k;
s22: calculating the mean square error between the global model estimated by the device k and the global model broadcasted by the base station when the base station downlink broadcasts
Figure BDA0002898677310000065
Calculating the mean square error between the model of base station aggregation and the target summation model when the model is uploaded by the equipment
Figure BDA0002898677310000071
Wherein
Figure BDA0002898677310000072
The variance of Gaussian white noise, | - | represents the absolute value of a scalar, and | | - | | represents the norm of a vector;
s23: introducing an auxiliary variable eta according to the mean square error limit of the base station and the mobile equipmentk≥0,ρk≥0,φk≥0,
Figure BDA0002898677310000073
Order to
Figure BDA0002898677310000074
Figure BDA0002898677310000075
Where γ is the base station allowed maximum mean square error bound, δkIs the maximum mean square error limit allowed by the mobile device k, I is the identity matrix with dimension N;
s24: based on maximum transmit power limits of the base station and the mobile device, the method comprises
Figure BDA0002898677310000076
And | | w | | non-conducting phosphor2≤PBS；
S25: alternately iterating and solving by using an interior point method or directly calling a CVX tool package
Figure BDA0002898677310000077
Get the weight value x ═ x of all mobile devices1,…,xK]T；
S26: determining target value | x | non-woven phosphor1Whether to converge or not, if the target value | | x | | non-woven phosphor1And when the convergence occurs, the mobile devices are reordered according to the priority selected by the current weight values x of all the mobile devices, otherwise, the step S22 is executed again.
S3: base station is reordered x according to selected priority of equipmentπ(1)≤…≤xπ(K)Performing feasibility detection to obtain the selected equipment set S in the t-th round of learningtPi (1), pi (2), …, pi (m) }, where m ∈ [1, K]Is the maximum value that makes the device set feasible, and then the transmit beam w, receive beam z of the base station and the receiver v of the selected mobile device k are designedkAnd a transmitter bk。
It should be noted that the serial numbers K of the K mobile devices are 1, …, and K, respectively, and when the mobile devices are prioritized according to the weight values x of all current mobile devices, the weights x of the K mobile devices need to be reordered first1,…,xKAccording to the weightThe values are sorted from small to large, then the serial number K of the 1 st mobile device in the sequence is assigned pi (1), the serial number K of the 2 nd mobile device in the sequence is assigned pi (2), and so on, and the serial number K of the last 1 mobile device in the sequence is assigned pi (K). Thus, a sequence reordered by a selected priority for a mobile device can be represented as xπ(1)≤…≤xπ(K)That is, the priority of the mobile device with sequence number pi (1) is the highest, and the priority of the mobile device with sequence number pi (K) is the lowest.
In this step, the selected set of devices S is obtained by a feasibility testtDesigning a base station transmitting beam w, a receiving beam z and a receiver v of a selected mobile device kkAnd a transmitter bkThe method comprises the following steps:
s31: the initialization parameter m is equal to K,
Figure BDA0002898677310000081
vk＝1，
Figure BDA0002898677310000082
z＝[1,0,…,0]T；
s32: reordering x according to device selected priorityπ(1)≤…≤xπ(K)The selected device set is St＝{π(1),π(2),…,π(m)}；
S33: when the downlink broadcast of the base station is calculated, the selected mobile equipment k belongs to StMean square error between estimated global model and global model broadcast by base station
Figure BDA0002898677310000083
Calculating the mean square error between the model aggregated by the base station and the target summation model when the selected mobile equipment uploads the local model
Figure BDA0002898677310000084
S34: according to the base station and the selected mobile equipment k belonging to StMean square error limitation of, introducing an auxiliary variable ηk≥0,ρk≥0,φk≥0,
Figure BDA0002898677310000085
Order to
Figure BDA0002898677310000086
Figure BDA0002898677310000087
S35: according to the base station and the selected mobile equipment k ∈ StMaximum transmit power limit of
Figure BDA0002898677310000088
And | | w | | non-conducting phosphor2≤PBS；
S36: alternately iterating and solving by using an interior point method or directly calling a CVX tool package
Figure BDA0002898677310000089
The solution with the smallest value of (c); if the solution exists, detecting the value of the mean square error of the base station, and if the solution satisfies the condition
Figure BDA00028986773100000810
To obtain w, vk,bkAnd z, and outputting the value of m, the selected device is St-pi (1), pi (2), …, pi (m) }; if no solution or not satisfied
Figure BDA0002898677310000091
Let m be m-1 and the execution start again from step S32.
S4: base station through downlink via transmitting beam w to selected device set StBroadcasting the updated global model q[t -1]。
S5: the selected mobile device k belongs to StBy receivers vkRecovering a received global model
Figure BDA0002898677310000092
And training local data set DkObtaining an updated local model
Figure BDA0002898677310000093
Then via transmitter bkAnd simultaneously transmitted to the base station via the uplink.
S6: after receiving the signals transmitted by the mobile equipment, the base station recovers the aggregated local model through the receiving beam z, and calculates the weighted average of the local models of all the mobile equipment as an updated global model q[t]。
S7: judging a global model q[t]Whether the global model is converged or not, if the global model is converged, the global model q is output[t]Otherwise, let t be t +1, the execution is resumed from step S2.
In this step, the method for updating the local model includes: training local data set D using convolutional neural networkkThe gradient updating method in the network is a random gradient descent method, and the local model updating formula is
Figure BDA0002898677310000094
Wherein
Figure BDA0002898677310000095
Is a gradient operation, Lk(. cndot.) is a loss function, and θ is the learning rate.
Computer simulation shows that, as shown in fig. 2, in the large-scale access method for the edge-oriented intelligent network provided by the invention, the larger the channel estimation error is, the fewer the selected devices are, but as the number of antennas increases, the performance can be obviously improved. Fig. 3 shows that the performance of the robust access method considering channel errors provided by the present invention is close to the ideal case and far superior to the non-robust access method. Therefore, the invention provides an effective access method for the edge intelligent network with large-scale equipment.

Claims (5)

1. A large-scale access method facing to an edge intelligent network is characterized by comprising the following steps:
s1: a base station with N antennas is deployed in advance, and a mobile device with K single antennas passes through the base stationAccessing a wireless network for federal learning, and regarding the global model as a round of learning when updating once; channel state information from the base station to the kth mobile device is obtained through channel estimation or feedback
Figure FDA0002898677300000011
The sequence number K is 1, …, K, and the real channel state information is
Figure FDA0002898677300000012
Wherein ekFor estimating the error vector for the channel, the norm of which has a boundary value epsilonkSatisfy | | | ek||≤εk(ii) a Initializing the number t of learning rounds to be 1;
s2: in the t-th round of learning, the base station performs ascending sequencing on all the mobile devices based on the channel state information to obtain a selected priority reordering xπ(1)≤…≤xπ(K)Wherein with xk∈[0,1]Representing weights of a kth mobile device, said selected priority reordering being according to xkIs ordered by the value size of xkThe smaller the value of (A) is, the higher the priority is selected, and the higher the ranking is; pi (1) is a serial number corresponding to the mobile equipment with the minimum weight value, pi (i) is a serial number corresponding to the ith mobile equipment which is ranked in the front after the weights of K mobile equipment are ranked from small to large, and pi (K) is a serial number corresponding to the mobile equipment with the maximum weight value; the values of pi (1), … and pi (K) are respectively one of 1, … and K and are not repeated;
s3: base station is reordered x according to selected priority of equipmentπ(1)≤…≤xπ(K)Performing feasibility detection to obtain the selected equipment set S in the t-th round of learningtPi (1), pi (2), …, pi (m) }, where m ∈ [1, K]Is the maximum value that makes the device set feasible, and then the transmit beam w, receive beam z of the base station and the receiver v of the selected mobile device k are designedkAnd a transmitter bk；
S4: base station through downlink via transmitting beam w to selected device set StBroadcasting the updated global model q[t-1]；
S5: is selected byThe mobile device k in (E) StBy receivers vkRecovering a received global model
Figure FDA0002898677300000013
And training local data set DkObtaining an updated local model
Figure FDA0002898677300000014
Then via transmitter bkSimultaneously transmitting to the base station through the uplink;
s6: after receiving the signals transmitted by the mobile equipment, the base station recovers the aggregated local model through the receiving beam z, and calculates the weighted average of the local models of all the mobile equipment as an updated global model q[t]；
S7: judging a global model q[t]Whether the global model is converged or not, if the global model is converged, the global model q is output[t]Otherwise, let t be t +1, the execution is resumed from step S2.
2. The large-scale access method for the edge-oriented intelligent network according to claim 1, wherein in step S2, the method for the base station to rank all the mobile devices based on the channel state information to obtain the selected priority ranking comprises:
s21: transmitting wave beam when initializing base station broadcasting global model
Figure FDA0002898677300000021
Receiver v of mobile device kkTransmitter when mobile device k uploads local model 1
Figure FDA0002898677300000022
Base station receiving beam z ═ 1,0, …,0]TIn which P isBSFor the maximum transmission power, P, of the base stationkIs the maximum transmit power of mobile device k;
s22: when the downlink broadcast of the base station is calculated, the global model estimated by the device k and the global model broadcast by the base stationMean square error of
Figure FDA0002898677300000023
Calculating the mean square error between the model of base station aggregation and the target summation model when the model is uploaded by the equipment
Figure FDA0002898677300000024
Wherein
Figure FDA0002898677300000025
The variance of Gaussian white noise, | - | represents the absolute value of a scalar, and | | - | | represents the norm of a vector;
s23: introducing an auxiliary variable eta according to the mean square error limit of the base station and the mobile equipmentk≥0,ρk≥0,φk≥0,
Figure FDA0002898677300000026
Order to
Figure FDA0002898677300000027
Figure FDA0002898677300000028
Where γ is the base station allowed maximum mean square error bound, δkIs the maximum mean square error limit allowed by the mobile device k, I is the identity matrix with dimension N;
s24: based on maximum transmit power limits of the base station and the mobile device, the method comprises
Figure FDA0002898677300000029
And | | w | | non-conducting phosphor2≤PBS；
S25: solve by alternating iterations
Figure FDA0002898677300000031
Get the weight value x ═ x of all mobile devices1,…,xK]T；
S26: judgmentNon-target value | | x | | non-conducting phosphor1Whether to converge or not, if the target value | | x | | non-woven phosphor1And when the convergence occurs, the mobile devices are reordered according to the priority selected by the current weight values x of all the mobile devices, otherwise, the step S22 is executed again.
3. The large-scale access method for edge-oriented intelligent network as claimed in claim 1, wherein in step S3, the selected device set S is obtained through feasibility detectiontDesigning a base station transmitting beam w, a receiving beam z and a receiver v of a selected mobile device kkAnd a transmitter bkThe method comprises the following steps:
s31: the initialization parameter m is equal to K,
Figure FDA0002898677300000032
vk＝1，
Figure FDA0002898677300000033
z＝[1,0,…,0]T；
s32: reordering x according to device selected priorityπ(1)≤…≤xπ(K)The selected device set is St＝{π(1),π(2),…,π(m)}；
S33: when the downlink broadcast of the base station is calculated, the selected mobile equipment k belongs to StMean square error between estimated global model and global model broadcast by base station
Figure FDA0002898677300000034
Calculating the mean square error between the model aggregated by the base station and the target summation model when the selected mobile equipment uploads the local model
Figure FDA0002898677300000035
S34: according to the base station and the selected mobile equipment k belonging to StMean square error limitation of, introducing an auxiliary variable ηk≥0,ρk≥0,φk≥0,
Figure FDA0002898677300000036
Order to
Figure FDA0002898677300000037
Figure FDA0002898677300000038
S35: according to the base station and the selected mobile equipment k ∈ StMaximum transmit power limit of
Figure FDA0002898677300000039
And | | w | | non-conducting phosphor2≤PBS；
S36: solve by alternating iterations
Figure FDA0002898677300000041
The solution with the smallest value of (c); if the solution exists, detecting the value of the mean square error of the base station, and if the solution satisfies the condition
Figure FDA0002898677300000042
To obtain w, vk,bkAnd z, and outputting the value of m, the selected device is St-pi (1), pi (2), …, pi (m) }; if no solution or not satisfied
Figure FDA0002898677300000043
Let m be m-1 and the execution start again from step S32.
4. The large-scale access method for the edge-oriented intelligent network according to claim 1, wherein in step S5, the method for updating the local model includes: training local data set D using convolutional neural networkkThe gradient updating method in the network is a random gradient descent method, and the local model updating formula is
Figure FDA0002898677300000044
Wherein
Figure FDA0002898677300000045
Is a gradient operation, Lk(. cndot.) is a loss function, and θ is the learning rate.
5. The large-scale access method for the edge-oriented intelligent network as claimed in claim 2 or 3, wherein in S25 and S36, the alternate iterative solution is performed by using an interior point method or directly calling a CVX toolkit.
CN202110049384.6A 2021-01-14 2021-01-14 Large-scale access method for edge-oriented intelligent network Active CN112911608B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110049384.6A CN112911608B (en) 2021-01-14 2021-01-14 Large-scale access method for edge-oriented intelligent network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110049384.6A CN112911608B (en) 2021-01-14 2021-01-14 Large-scale access method for edge-oriented intelligent network

Publications (2)

Publication Number Publication Date
CN112911608A true CN112911608A (en) 2021-06-04
CN112911608B CN112911608B (en) 2021-11-23

Family

ID=76114245

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110049384.6A Active CN112911608B (en) 2021-01-14 2021-01-14 Large-scale access method for edge-oriented intelligent network

Country Status (1)

Country Link
CN (1) CN112911608B (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113965281A (en) * 2021-10-08 2022-01-21 上海科技大学 Air computing system design method based on end-to-end learning
CN114189899A (en) * 2021-12-10 2022-03-15 东南大学 User equipment selection method based on random aggregation beam forming
CN114204971A (en) * 2021-12-10 2022-03-18 东南大学 Iterative aggregation beamforming design and user equipment selection method
WO2023082207A1 (en) * 2021-11-12 2023-05-19 Lenovo (Beijing) Limited Methods and apparatuses for user equipment selecting and scheduling in intelligent wireless system
CN116781518A (en) * 2023-08-23 2023-09-19 北京光函数科技有限公司 Federal multi-arm slot machine learning method and system

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2006125456A1 (en) * 2005-05-24 2006-11-30 Telecom Italia, S.P.A. Method for dimensioning a data packets handler apparatus in a packet-switched mobile communications network
US20100234071A1 (en) * 2009-03-12 2010-09-16 Comsys Communication & Signal Processing Ltd. Vehicle integrated communications system
CN103703814A (en) * 2011-07-28 2014-04-02 瑞典爱立信有限公司 Beamforming for cell edge capacity improvement in a heterogeneous network
CN106953709A (en) * 2017-03-20 2017-07-14 浙江大学 A kind of descending nonopiate multi-access method of multiaerial system based on spatial modulation
CN109039410A (en) * 2018-06-27 2018-12-18 南京邮电大学 A kind of beam-forming method of isomery cloud Radio Access Network and the communication network for realizing this method
CN109788074A (en) * 2018-03-19 2019-05-21 南京邮电大学 A kind of Edge intelligence service system
CN111369042A (en) * 2020-02-27 2020-07-03 山东大学 Wireless service flow prediction method based on weighted federal learning
CN111756848A (en) * 2020-06-28 2020-10-09 河海大学 QoS optimization method based on federal learning and mobile perception under mobile edge environment
CN111866869A (en) * 2020-07-07 2020-10-30 兰州交通大学 Federal learning indoor positioning privacy protection method facing edge calculation
CN112153650A (en) * 2020-08-31 2020-12-29 西北农林科技大学 Reliable federal learning method and system based on terminal reputation in wireless network
CN112181666A (en) * 2020-10-26 2021-01-05 华侨大学 Method, system, equipment and readable storage medium for equipment evaluation and federal learning importance aggregation based on edge intelligence

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2006125456A1 (en) * 2005-05-24 2006-11-30 Telecom Italia, S.P.A. Method for dimensioning a data packets handler apparatus in a packet-switched mobile communications network
US20100234071A1 (en) * 2009-03-12 2010-09-16 Comsys Communication & Signal Processing Ltd. Vehicle integrated communications system
CN103703814A (en) * 2011-07-28 2014-04-02 瑞典爱立信有限公司 Beamforming for cell edge capacity improvement in a heterogeneous network
CN106953709A (en) * 2017-03-20 2017-07-14 浙江大学 A kind of descending nonopiate multi-access method of multiaerial system based on spatial modulation
CN109788074A (en) * 2018-03-19 2019-05-21 南京邮电大学 A kind of Edge intelligence service system
CN109039410A (en) * 2018-06-27 2018-12-18 南京邮电大学 A kind of beam-forming method of isomery cloud Radio Access Network and the communication network for realizing this method
CN111369042A (en) * 2020-02-27 2020-07-03 山东大学 Wireless service flow prediction method based on weighted federal learning
CN111756848A (en) * 2020-06-28 2020-10-09 河海大学 QoS optimization method based on federal learning and mobile perception under mobile edge environment
CN111866869A (en) * 2020-07-07 2020-10-30 兰州交通大学 Federal learning indoor positioning privacy protection method facing edge calculation
CN112153650A (en) * 2020-08-31 2020-12-29 西北农林科技大学 Reliable federal learning method and system based on terminal reputation in wireless network
CN112181666A (en) * 2020-10-26 2021-01-05 华侨大学 Method, system, equipment and readable storage medium for equipment evaluation and federal learning importance aggregation based on edge intelligence

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
JING LIU: "Performance Evaluation of Integrated Multi-Access Edge Computing and Fiber-Wireless Access Networks", 《 IEEE ACCESS ( VOLUME: 6)》 *
ZHENG CAI: "Task Assignment with Precedence Constraint over Networks: A Case Study of Computation-Communication Convergence", 《2018 10TH INTERNATIONAL CONFERENCE ON WIRELESS COMMUNICATIONS AND SIGNAL PROCESSING (WCSP)》 *
张启宁: "基于联邦学习和循环神经网络的无线网络入侵检测研究", 《中国计算机用户协会网络应用分会2020年第二十四届网络新技术与应用年会论文集》 *
王慧强: "认知网络基于跨层感知的接入选择机制", 《通信学报》 *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113965281A (en) * 2021-10-08 2022-01-21 上海科技大学 Air computing system design method based on end-to-end learning
CN113965281B (en) * 2021-10-08 2023-09-22 上海科技大学 Air computing system design method based on end-to-end learning
WO2023082207A1 (en) * 2021-11-12 2023-05-19 Lenovo (Beijing) Limited Methods and apparatuses for user equipment selecting and scheduling in intelligent wireless system
CN114189899A (en) * 2021-12-10 2022-03-15 东南大学 User equipment selection method based on random aggregation beam forming
CN114204971A (en) * 2021-12-10 2022-03-18 东南大学 Iterative aggregation beamforming design and user equipment selection method
CN114189899B (en) * 2021-12-10 2023-03-31 东南大学 User equipment selection method based on random aggregation beam forming
CN114204971B (en) * 2021-12-10 2024-01-30 东南大学 Iterative aggregate beam forming design and user equipment selection method
CN116781518A (en) * 2023-08-23 2023-09-19 北京光函数科技有限公司 Federal multi-arm slot machine learning method and system
CN116781518B (en) * 2023-08-23 2023-10-24 北京光函数科技有限公司 Federal multi-arm slot machine learning method and system

Also Published As

Publication number Publication date
CN112911608B (en) 2021-11-23

Similar Documents

Publication Publication Date Title
CN112911608B (en) Large-scale access method for edge-oriented intelligent network
CN110855585B (en) Channel estimation method of millimeter wave large-scale MIMO system
CN109474388B (en) Low-complexity MIMO-NOMA system signal detection method based on improved gradient projection method
CN110167176B (en) Wireless network resource allocation method based on distributed machine learning
CN110233653B (en) Blind multipath identification method and system of MIMO system based on weighted integration clustering algorithm
CN108768472B (en) Near space high altitude platform space-polarization diversity MIMO channel modeling method
CN111835406A (en) Robust precoding method suitable for energy efficiency and spectral efficiency balance of multi-beam satellite communication
US11742901B2 (en) Deep learning based beamforming method and apparatus
CN115632684B (en) Transmission strategy design method of perception and communication integrated system
CN114337976A (en) Transmission method combining AP selection and pilot frequency allocation
CN114219354A (en) Resource allocation optimization method and system based on federal learning
CN115767758A (en) Equipment scheduling method based on combination of channel and local model update
CN106788631A (en) A kind of extensive MIMO reciprocities calibration method based on local alignment
Aygün et al. Over-the-air federated learning with energy harvesting devices
CN114268348A (en) Honeycomb-free large-scale MIMO power distribution method based on deep reinforcement learning
CN114204971A (en) Iterative aggregation beamforming design and user equipment selection method
CN108566227B (en) Multi-user detection method
CN114764610A (en) Channel estimation method based on neural network and communication device
CN116170066B (en) Load prediction method for low-orbit satellite Internet of things
CN107346985B (en) Interference alignment method combined with transmitting antenna selection technology
CN114501353B (en) Communication information sending and receiving method and communication equipment
CN113056029B (en) Energy-limited terminal random access method suitable for low-earth-orbit satellite network
Wang et al. Integrated Sensing, Communication, and Computation Over-the-Air: Beampattern Design
CN114710187A (en) Power distribution method for multi-cell large-scale MIMO intelligent communication under dynamic user number change scene
CN115843045A (en) Data acquisition method and device

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