CN117455014A - Federal learning method and device based on low orbit satellite constellation - Google Patents
Federal learning method and device based on low orbit satellite constellation Download PDFInfo
- Publication number
- CN117455014A CN117455014A CN202311533269.1A CN202311533269A CN117455014A CN 117455014 A CN117455014 A CN 117455014A CN 202311533269 A CN202311533269 A CN 202311533269A CN 117455014 A CN117455014 A CN 117455014A
- Authority
- CN
- China
- Prior art keywords
- orbit
- low
- satellite
- communication
- model
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 92
- 238000004891 communication Methods 0.000 claims abstract description 143
- 230000005540 biological transmission Effects 0.000 claims abstract description 70
- 239000012634 fragment Substances 0.000 claims description 154
- 238000006467 substitution reaction Methods 0.000 claims description 17
- 238000012545 processing Methods 0.000 claims description 15
- 238000006116 polymerization reaction Methods 0.000 claims description 14
- 238000004590 computer program Methods 0.000 claims description 13
- 238000003860 storage Methods 0.000 claims description 9
- 230000000379 polymerizing effect Effects 0.000 claims description 8
- 230000002776 aggregation Effects 0.000 description 30
- 238000004220 aggregation Methods 0.000 description 30
- 230000008569 process Effects 0.000 description 29
- 238000012549 training Methods 0.000 description 8
- 230000001360 synchronised effect Effects 0.000 description 7
- 238000010801 machine learning Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000013461 design Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000009365 direct transmission Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000009987 spinning Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16Y—INFORMATION AND COMMUNICATION TECHNOLOGY SPECIALLY ADAPTED FOR THE INTERNET OF THINGS [IoT]
- G16Y10/00—Economic sectors
- G16Y10/75—Information technology; Communication
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Radio Relay Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Artificial Intelligence (AREA)
- Business, Economics & Management (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Economics (AREA)
- General Business, Economics & Management (AREA)
- General Engineering & Computer Science (AREA)
Abstract
The invention provides a federal learning method and a federal learning device based on a low-orbit satellite constellation, and relates to the technical field of satellite Internet of things, wherein the method comprises the following steps: under the condition that the orbit can realize annular communication and a converging satellite exists in the orbit, after local model updating is carried out, transmitting model updating parameters of the orbit to the converging satellite so that the converging satellite can transmit model updating parameters of all low-orbit satellites in the orbit to a ground station; the convergence satellite is a low-orbit satellite in which the communication time with the ground station is longer than the sum of the convergence time and the transmission time in the orbit; the convergence time is the time required for converging the model update parameters of all the low-orbit satellites in the orbit; the transmission time is the time required to transmit the aggregated model update parameters to the ground station. The invention realizes two-layer federation learning based on synchronization and asynchronization, improves the time efficiency of the federation learning algorithm in the low-orbit satellite constellation, and reduces the complexity of the federation learning algorithm.
Description
Technical Field
The invention relates to the technical field of satellite Internet of things, in particular to a federal learning method and device based on low-orbit satellite constellation.
Background
The low orbit satellite provides automatic storage and forwarding data communication service in the Internet of things network, valuable data are accumulated in the satellite while the Internet of things task is completed, and the data can be used for obtaining an intelligent model by utilizing a machine learning or deep learning technology, so that the application efficiency of the Internet of things is further improved. However, the problems of data privacy of the internet of things and communication overhead of the native data pose challenges for the process of multi-satellite continuous direct transmission of data to the central server. Federal learning thus provides a viable solution to the above-described problem by using local model parameters to train a machine learning model rather than sharing data sets on satellites.
However, due to the scene characteristics of the low-orbit satellites, firstly, the contact time between each satellite serving as a client and the ground station is too short to wait for the distribution of global model parameter update after the uploading of local model parameters, and the next contact is usually needed for obtaining a new global model. In addition, the invisible time between each satellite and the ground station is a long period and cannot be added to federal learning. Therefore, the convergence rate of federal learning through satellite constellations can be very slow. When the multi-orbit multi-satellite participates in the learning algorithm, the scheduling and calculating process of the ground station is more complex.
Some current studies suggest that predictability of satellite motion in constellation scenarios helps federal learning algorithm design, which enables synchronized federal learning through ground station scheduling and inter-satellite links when the training process is coordinated by the ground stations. However, due to earth spinning, there is a communication blockage with the ground station throughout the orbit, and the latency for one round of communication will be longer when multiple orbits and satellites are added.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a federal learning method and a federal learning device based on a low-orbit satellite constellation, which are used for improving the time efficiency of a federal learning algorithm in the low-orbit satellite constellation and reducing the complexity of the federal learning algorithm.
The invention provides a federal learning method based on a low-orbit satellite constellation, which comprises the following steps:
under the condition that the orbit can realize annular communication and a converging satellite exists in the orbit, after local model updating is carried out, transmitting model updating parameters of the converging satellite to the converging satellite so that the converging satellite can transmit model updating parameters of all low-orbit satellites in the orbit to a ground station;
wherein the converged satellite is a low-orbit satellite in which the communicable time with the ground station is longer than the sum of the converged time and the transmission time in the orbit; the convergence time is the time required for converging the model update parameters of all the low-orbit satellites in the orbit; the transmission time is the time required for transmitting the aggregated model update parameters to the ground station.
In some embodiments, the method further comprises:
and under the condition that the orbit can realize annular communication and all the low-orbit satellites in the orbit can not communicate with the ground station, after local model updating is carried out, transmitting own model updating parameters to the low-orbit satellites in the next communication according to the communication direction of the annular communication, and receiving the model updating parameters transmitted by the low-orbit satellites in the last communication until each low-orbit satellite in the orbit holds the model updating parameters of all the low-orbit satellites.
In some embodiments, the transmitting the model update parameter to the low-orbit satellite in the next communication according to the communication direction of the ring communication and receiving the model update parameter transmitted by the low-orbit satellite in the last communication include:
splitting the update parameters of the self model into p segments;
in the ith round of transmission, fragments are transmittedTransmitting to said low-orbit satellite in the next communication and receiving the fragment of the low-orbit satellite transmission in the last communication>
Wherein the saidUpdating parameters for the model of the xth segment of its own p segments, where x is p-k- (i [ mod ] ]p) +2, said k being the ordering of itself in said track; said->Updating parameters for a model of a y-th segment of the p segments of the last-communicated low-orbit satellite, wherein y is x+1[ mod ]]p, the k-1 is the order of the low-orbit satellite in the last communication in the orbit.
In some embodiments, the method further comprises:
the fragments are subjected toIs transmitted to the low-orbit satellite in the next communication for the low-orbit satellite in the next communication to be associated with the fragment +.>Fragments of the same index, identical to said fragments->Polymerizing;
receiving the segment of the low-orbit satellite transmission in the last communicationIndex of (2);
will be associated with the fragmentFragments of the same index, identical to said fragments->Polymerization is carried out.
In some embodiments, the method further comprises:
the fragments are subjected toTransmitting the corresponding satellite index to the low-orbit satellite in the next communication for the low-orbit satellite in the next communication, wherein the low-orbit satellite is in the same way as the fragment ∈ ->Satellite indexes corresponding to the same-index fragments are contained in the fragments +.>In the case of the corresponding satellite index, will be +_with the fragment>Substitution of the same indexed fragment with said fragment +. >
Receiving the segment of the low-orbit satellite transmission in the last communicationA corresponding satellite index;
at the same time as the fragmentSatellite indexes corresponding to the same-index fragments are contained in the fragments +.>In the case of the corresponding satellite index, will be +_with the fragment>Substitution of the same indexed fragment with said fragment +.>
In some embodiments, the method further comprises:
and under the condition that the track can not realize annular communication and can communicate with the ground station, transmitting the model updating parameters to the ground station after carrying out local model updating.
In some embodiments, the method further comprises:
a local model update is performed upon determining that the in-orbit cannot achieve annular communication and that all low-orbit satellites in the in-orbit cannot communicate with the ground station.
The invention also provides a federal learning device based on the low-orbit satellite constellation, which comprises:
the first processing module is used for transmitting the model updating parameters of the first processing module to the converging satellites after carrying out local model updating under the condition that the orbit can realize annular communication and the converging satellites exist in the orbit, so that the converging satellites can transmit the model updating parameters of all low-orbit satellites in the orbit to the ground station;
Wherein the converged satellite is a low-orbit satellite in which the communicable time with the ground station is longer than the sum of the converged time and the transmission time in the orbit; the convergence time is the time required for converging the model update parameters of all the low-orbit satellites in the orbit; the transmission time is the time required for transmitting the aggregated model update parameters to the ground station.
The invention also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the federal learning method based on low-orbit satellite constellation as described in any one of the above when executing the program.
The present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements a federal learning method based on a low-orbit satellite constellation as described in any of the above.
The invention also provides a computer program product comprising a computer program which when executed by a processor implements a federal learning method based on a low-earth satellite constellation as defined in any one of the above.
According to the federal learning method and device based on the low-orbit satellite constellation, the mode that the original low-orbit satellite directly transmits the model updating parameters to the ground station is changed into the mode that the model updating parameters of all low-orbit satellites in orbit are converged in the converging satellite, and then all the model updating parameters are transmitted to the ground station by the converging satellite, so that intra-orbit aggregation is increased on the basis of local model updating and global aggregation, two-layer federal learning based on synchronous federal learning and asynchronous federal learning is realized, the time efficiency of a federal learning algorithm in the low-orbit satellite constellation is improved, and the complexity of the federal learning algorithm is reduced.
Drawings
In order to more clearly illustrate the invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.
Fig. 1 is a schematic view of a scenario of a low orbit satellite constellation internet of things;
fig. 2 is a schematic flow chart of a federal learning method based on low-orbit satellite constellation provided by the invention;
fig. 3 is a schematic structural diagram of a federal learning device based on low-orbit satellite constellation according to the present invention;
fig. 4 is a schematic structural diagram of an electronic device provided by the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, not all embodiments. The embodiments of the present invention and the features in the embodiments may be combined with each other without collision. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.
It is further intended that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.
The term "at least one" in the present invention means one or more, and "a plurality" means two or more. The terms "first," "second," "third," "fourth," and the like in this disclosure, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order.
In embodiments of the invention, words such as "exemplary" or "such as" are used to mean serving as an example, instance, or illustration. Any embodiment or design described herein as "exemplary" or "e.g." in an embodiment should not be taken as preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "such as" is intended to present related concepts in a concrete fashion.
In order to facilitate a clearer understanding of the embodiments of the present invention, some related technical knowledge will be described first.
(1) Federal study
In conventional machine learning, data typically needs to be stored centrally on a central server, and then a model is trained on the central server. Federal learning is a decentralized machine learning method aimed at training a global model without transmitting raw data from the device to a central server.
In federal learning, data remains local to each device (e.g., mobile device, sensor, etc.), the model training process is performed on each device, and only model parameters are transmitted to a central server for aggregation. The core idea of federal learning is to implement data privacy protection and decentralized computing. Because the data does not leave the local equipment, the privacy of the user is protected, and direct communication is not needed between different equipment, so that the data transmission quantity is reduced. This is very useful for scenarios where data privacy sensitivity is high, bandwidth is limited, or connection is unstable.
A federal learning system is typically comprised of a server and a plurality of devices. The convergence to the global model can be achieved with a small resource consumption over a certain number of iterations.
The basic flow of federal learning is as follows: 1) Initializing a global model: the central server initializes a global model that will be trained on the various devices. 2) Training equipment: each device trains the global model using local data. The training process may use various machine learning algorithms, such as neural networks, decision trees, and the like. 3) Updating a local model: after training is completed on the device, the device only transmits the updated parameters (gradients) of the model to the central server, rather than transmitting the raw data. 4) Central server aggregation: the central server gathers model update parameters from different devices and aggregates them to generate a new version of the global model. 5) Updating the device model: the new version of the global model is transmitted back to the device, which takes it as the new initial model and goes to the next round of training.
This process may be performed in multiple rounds, each of which gradually improves the global model. Federal learning can be applied in many fields, such as personalized models on mobile devices, analysis of medical data, collaboration of internet of things devices, and the like.
The most widely used conventional synchronous federal learning and asynchronous federal learning are proposed sequentially. The traditional synchronous federal learning process is as follows: the central server broadcasts the latest model to the distributed devices, the devices train the model on the local data set, and upload the local model parameters to the central server. After all clients complete the upload task, the central server aggregates the model parameters to update the global model, in other words, the central server cannot update the global model until all planned devices have completed the upload parameters at this time. This constitutes an iteration of the global model. However, due to the hysteresis problem caused by sparse task activation and sparse communications, the speed of synchronous federal learning is typically slow.
To overcome these challenges, asynchronous federal learning has been proposed to take advantage of asynchronous training. In asynchronous federal learning, the device also trains on local data according to the latest global model received and uploads local model parameters to the central server. However, once the central server receives the model parameters from any device, the central server may immediately update the global model by weighted averaging. Thus, in asynchronous federal learning, the central server and devices are updated asynchronously. Asynchronous federal learning is not hindered as long as the communication between the central server and one of the devices can meet the conditions for parameter upload and parameter broadcast.
(2) Communication of low orbit satellite constellation internet of things
Fig. 1 is a schematic view of a scenario of the internet of things of a low-orbit satellite constellation, and referring to fig. 1, the low-orbit satellite constellation includes a plurality of orbits, each orbit has a plurality of low-orbit satellites, each low-orbit satellite has a unique identifier, the low-orbit satellites can communicate with each other through an inter-satellite link, and the low-orbit satellites can communicate with a ground station through a satellite-ground link.
In the event that the distance between two low-orbit satellites is less than or equal to the distance threshold, then communication between the two low-orbit satellites may be via an inter-satellite link.
Let the kth low-orbit satellite on orbit m be S k,m The first low-orbit satellite in orbit m is marked as S l,m Low orbit satellite S k,m And low orbit satellite S l,m The expression of the distance threshold between is as follows:
wherein d th (l, k) denotes a low-orbit satellite S k,m And low-orbit satellite S l,m Distance threshold between, h l Representing a low-orbit satellite S l,m Height, r E Represents the earth radius, h k Representing a low-orbit satellite S k,m Is a high level of (2).
In low orbit satellite S k,m Elevation angle with ground station is greater than elevation angle threshold alpha th In the case of (i), i.eIn the case of (a), a low-orbit satellite S k,m May communicate with the ground station. Wherein,indicating the position of the ground station- >Representing a low-orbit satellite S k,m Is a position of (c).
With the sender denoted S and the receiver denoted R, the sender may be any low-orbit satellite in orbit and the receiver may be a neighboring low-orbit satellite or ground station. The expression of the delay between the sender and the receiver is as follows:
wherein T is SR Representing the delay between the sender and the receiver,representing the transmission delay between sender and receiver, < >>Representing the propagation delay between the sender and the receiver.
Wherein propagation delay between sender and receiverThe expression of (2) is as follows:
in the method, in the process of the invention,representing propagation delay between sender and receiver, L SR Representing the physical link length between sender and receiver, V SR Representing the propagation velocity in the wireless medium.
Wherein the transmission delay between sender and receiverThe expression of (2) is as follows:
in the method, in the process of the invention,representing the transmission delay between sender and receiver, M SR Representing the amount of data transferred between a sender and a receiver, R SR Representing the transmission data rate between the sender and the receiver.
Wherein the expression of the transmission data rate between the sender and the receiver is as follows:
wherein R is SR Representing the transmission data rate between sender and receiver, B SR Representing bandwidth, SNR between sender and receiver SR Representing the signal-to-noise ratio between the sender and the receiver.
Wherein, the expression of the signal-to-noise ratio between the sender and the receiver is as follows:
SNR SR ＝p SR |h SR |
in the SNR SR Representing the signal-to-noise ratio between sender and receiver, p SR Representing the signal power between sender and receiver, h SR Representing the channel state between the sender and the receiver.
Wherein the expression of the propagation delay between the sender and the receiver is as follows:
in the method, in the process of the invention,representing propagation delay between sender and receiver, L SR Representing the physical link length between sender and receiver, V SR Representing the propagation velocity in the wireless medium.
Fig. 2 is a flow chart of a federal learning method based on a low-orbit satellite constellation, and referring to fig. 2, the invention provides a federal learning method based on a low-orbit satellite constellation, the execution subject is a low-orbit satellite, and the method includes:
and under the condition that the orbit can form annular communication and the convergence satellite exists in the orbit, transmitting the model updating parameters of the convergence satellite to the convergence satellite after carrying out local model updating, so that the convergence satellite can transmit the model updating parameters of all the low-orbit satellites in the orbit to the ground station.
The convergence satellite is a low-orbit satellite which is in an orbit and has a communication time with the ground station longer than the sum of the convergence time and the transmission time; the convergence time is the time required for converging the model update parameters of all the low-orbit satellites in the orbit; the transmission time is the time required to transmit the aggregated model update parameters to the ground station.
In particular, the traditional federal learning architecture is low-earth satellites and ground stations, between which orbits are introduced in order to make full use of the inter-satellite links.
When all low-orbit satellites in one orbit can communicate with adjacent low-orbit satellites through inter-satellite links, i.e. when the orbit can realize annular communication, then the model update parameters can be aggregated in orbit.
When only a part of low-orbit satellites in one orbit can communicate with adjacent low-orbit satellites through inter-satellite links, namely, when the orbit cannot realize annular communication, each low-orbit satellite is used as a separate client, and the model updating parameters cannot be aggregated in the orbit.
The state of the orbit of the low-orbit satellite is recorded as S nei In S nei When=1, it indicates that the orbit of the low-orbit satellite can realize annular communication; at S nei When=0, it indicates that the low-orbit satellite is in an orbit in which the ring communication cannot be realized.
The model update parameters are finally transmitted to the ground station, and the communication state between the low-orbit satellite and the ground station is recorded as S g . At S g When=1, it indicates that the low-orbit satellite itself can communicate with the ground station, or that a communicable satellite having the same orbit can communicate with the ground station; at S g When=0, it is indicated that the low-orbit satellite itself cannot communicate with the ground station, and that there is no communicable satellite in the same orbit as the ground station, i.e., all low-orbit satellites in the orbit cannot communicate with the ground station.
In the case where the low-orbit satellite is in orbit to enable annular communication and the co-orbit communicable satellite is in communication with the ground station, S nei =1 and S g Condition of =1In some cases, the federal learning process includes: local model update, in-orbit aggregation and global aggregation. The specific process is as follows:
and in the local model updating stage, each low-orbit satellite in the orbit carries out local model updating, and model updating parameters of the low-orbit satellite are respectively acquired.
Here by low-orbit satellites S k,m For example, a description will be given of a procedure in which a low-orbit satellite performs local model update.
Low orbit satellite S k,m The model is updated with the local data set. The expression of the loss function is as follows:
where f () represents the loss function of each sample, |D k,w I represents the size of the local dataset, j represents each sample index, w represents the model update parameters, F k,w (w) denotes a low-orbit satellite S k,m Is lost to the local model.
Low orbit satellite S k,m And obtaining the model updating parameters of the current times on the basis of the model updating parameters obtained in the previous time according to the loss function. The expression of the model update parameters is as follows:
where, η represents a parameter update step size,representing local model gradients, F k,w Representing a low-orbit satellite S k,m τ represents the number of model iterations on the track, +.>Representing a low-orbit satellite S k,m Model update parameters obtained at τ -1 th time, < ->Representing a low-orbit satellite S k,m And updating parameters of the model obtained at the τ time.
In-orbit polymerization stage, firstly, selecting convergent satellites from the same-orbit communicable satellites. The selected convergence satellite should satisfy: communication time T with ground station access Greater than convergence time T agg And a transmission time T trans And (3) summing. The convergence time is the time required for the model update parameters of all the low orbit satellites in the orbit where the convergence satellites are converged. The transmission time is the time required by the converged satellite to transmit the converged model update parameters to the ground station.
Convergence time T agg The expression of (2) is as follows:
wherein T is agg Represents the convergence time, K, required for converging the model update parameters of all the low-orbit satellites in the orbit m Representing the number of low-orbit satellites in m orbits, T SR Representing the delay between the sender and the receiver.
Transmission time T trans The expression of (2) is as follows:
T rans ＝2T SG
wherein T is rans Representing the transmission time, T, required to transmit the aggregated model update parameters to the ground station SG Representing the delay of converging satellites to a ground station.
After the convergence satellite is selected, the convergence satellite sends a broadcast message to the low-orbit satellite in the orbit, and the broadcast message carries information of the convergence satellite. For example, coordinate information of the converged satellites, or position information of the converged satellites in the orbit, etc. And after receiving the broadcast messages sent by the convergence satellites, the low-orbit satellites in the orbit respectively transmit model updating parameters of the low-orbit satellites to the convergence satellites.
The process of converging the model update parameters of all the low-orbit satellites in the orbit can be regarded as a process of performing synchronous federal learning in the orbit. The expression of the model update parameters after in-orbit aggregation is as follows:
wherein w is m Representing model update parameters, w, after in-orbit aggregation k,m Representing a low-orbit satellite S k,m Model update parameters, |D k,m I denotes a low-orbit satellite S k,m Is represented by the size, τ, of the local dataset of (a) and the number of model iterations, K m Represents the total number of low-orbit satellites in m orbits, and k represents the ordering of the low-orbit satellites in m orbits.
In the global aggregation stage, because of the motion of the earth and the satellite constellation in the low-orbit satellite constellation and the shielding of the earth, each orbit establishes time step synchronization of contact with the ground station, the ground station executes asynchronous federal learning, namely, the ground station immediately executes global aggregation once receiving the model updating parameters. The convergence satellite transmits all model update parameters to the ground station, which performs global convergence immediately upon standing on the ground.
The asynchronous federal learning process updates the global model using partial model update parameters as follows:
in the method, in the process of the invention,representing global model update parameters->Model update parameters, alpha, representing new uploads to ground stations τ Weights, τ, representing model update parameters t The number of iterations of the global model is represented, and the number of iterations of the model on the orbit is represented.
Wherein, model updating parametersWeight of number alpha τ The expression of (2) is as follows:
wherein alpha is τ Weights representing model update parameters, a represents freshness threshold, a >0, b represents the attenuation degree, b>0, alpha represents the initial weight, τ t The number of iterations of the global model is represented, and the number of iterations of the model on the orbit is represented.
Wherein, the expression of the initial weight alpha is as follows:
where α represents the initial weight, |D k,m I denotes a low-orbit satellite S k,m Is the size, K, of the local data set of (2) m Represents the total number of low-orbit satellites in M orbits, k represents the ordering of the low-orbit satellites in M orbits, M represents the total number of all orbits in the low-orbit satellite constellation, and M represents the mth orbit in the low-orbit satellite constellation.
According to the federal learning method based on the low-orbit satellite constellation, the mode that the original low-orbit satellite directly transmits model updating parameters to the ground station is changed into the mode that the model updating parameters of all low-orbit satellites in orbit are converged at the converging satellite, and then all model updating parameters are transmitted to the ground station through the converging satellite, so that intra-orbit aggregation is increased on the basis of local model updating and global aggregation, two-layer federal learning based on synchronous federal learning and asynchronous federal learning is realized, the time efficiency of federal learning algorithm in the low-orbit satellite constellation is improved, and the complexity of the federal learning algorithm is reduced.
In some embodiments, the federal learning method based on low-orbit satellite constellation provided in the embodiments of the present invention further includes:
and under the condition that the orbit can realize annular communication and all the low-orbit satellites in the orbit can not communicate with the ground station, after the local model is updated, transmitting the model updating parameters to the low-orbit satellites in the next communication according to the communication direction of the annular communication, and receiving the model updating parameters transmitted by the low-orbit satellites in the last communication until each low-orbit satellite in the orbit holds the model updating parameters of all the low-orbit satellites.
In particular, in the case where the low-orbit satellites are in orbits in which annular communication can be achieved, and all of the low-orbit satellites in orbit cannot communicate with the ground station, i.e., S nei =1 and S g In the case of=0, the federal learning process includes: local model update and in-orbit aggregation.
The process of local model update is as described above and will not be described in detail here. The important point is to introduce this stage of in-orbit polymerization, which is specifically as follows:
since there are no converging satellites in orbit, in-orbit aggregation cannot be performed by transmitting model update parameters of all low-orbit satellites in orbit to the converging satellites.
And because the orbit can realize annular communication, after the local model update is carried out on each low-orbit satellite in the orbit, the model update parameters of each low-orbit satellite in the orbit can be transmitted to the low-orbit satellite in the next communication according to the communication direction of the annular communication, and the model update parameters transmitted by the low-orbit satellite in the last communication are received, so that the multi-round transmission is carried out until each low-orbit satellite in the orbit holds the model update parameters of all the low-orbit satellites, and the in-orbit aggregation is realized.
According to the federal learning method based on the low-orbit satellite constellation, under the condition that the orbit can realize annular communication and all low-orbit satellites in the orbit can not communicate with the ground station, the model updating parameters of the low-orbit satellites are transmitted according to the communication direction of the annular communication, and the model updating parameters of other low-orbit satellite transmissions are received, so that the in-orbit aggregation under the condition of no aggregation satellite is realized.
In some embodiments, transmitting the model update parameter to the low-orbit satellite in the next communication and receiving the model update parameter transmitted by the low-orbit satellite in the previous communication according to the communication direction of the ring communication comprises:
Splitting the update parameters of the self model into p segments;
in the ith round of transmission, fragments are transmittedTransmitting to the low-orbit satellite in the next communication and receiving the fragment +.>
Wherein,updating parameters for the model of the xth segment of its own p segments, x being p-k- (i [ mod ]]p) +2, k is the order of itself in the track; />Updating parameters for the model of the y-th segment of the p segments of the low orbit satellite in the last communication, y being x+1[ mod ]]p, k-1 is the ordering of the low-orbit satellites in the orbit of the last communication.
Specifically, in order to reduce the data volume of a single transmission, each low-orbit satellite in the orbit splits its own model update parameters into p segments. Preferably, p is equal to the total number of low-orbit satellites in orbit. For example, if there are 3 low-orbit satellites in orbit, the model update parameters are split into 3 segments.
By low-orbit satellites S k,m For illustration, low-orbit satellite S k,m Model update parameter w of (2) k Updating the model to the parameter w k Splitting into p fragments, i.e
In the ith round of transmission, low-orbit satellite S k,m Fragments are processedTo the low orbit satellite in the next communication. Wherein,updating parameters for the model of the xth segment of its own p segments, x being p-k- (i [ mod ] ]p) +2, k is low-orbit satellite S k,m The ranking value in the orbit (ranking value can be understood as satellite index).
Low orbit satellite S k,m Receiving segments of low-orbit satellite transmissions in a previous communicationWherein (1)>Updating parameters for the model of the y-th segment of the p segments of the low orbit satellite in the last communication, y being x+1[ mod ]]p, k-1 is the ordering of the low-orbit satellites in the orbit of the last communication.
After 2 (p-1) passes, all low-orbit satellites in orbit hold model update parameters of the low-orbit satellites, and each low-orbit satellite transmitsWherein M is SR Representing the amount of data transferred between the sender and the receiver.
According to the federal learning method based on the low-orbit satellite constellation, provided by the embodiment of the invention, the model updating parameters are split into a plurality of fragments, and the fragments are used as units for transmission, so that the data volume of single transmission is reduced.
In some embodiments, the federal learning method based on low-orbit satellite constellation provided in the embodiments of the present invention further includes:
fragments are processedThe index at the next is transmittedA low-orbit satellite for communication, so that the low-orbit satellite for the next communication will be associated with the fragment +.>Fragments of the same index, and fragments->Polymerizing;
Receiving segments of low-orbit satellite transmissions in a previous communicationIndex of (2);
will be associated with the fragmentFragments of the same index, and fragments->Polymerization is carried out.
In particular, since the model update parameters of the low orbit satellite for the next communication are also split into p segments, the segments are then divided into two segmentsAfter transmission to the low-orbit satellite in the next communication, the low-orbit satellite in the next communication should be made clear that the fragment +.>With which fragment of itself it is best to aggregate.
Thus, in the process of dividing the segmentTransmitting to the low orbit satellite in the next communication, and simultaneously transmitting the fragment +.>Is transmitted to the low-orbit satellite in the next communication so that the low-orbit satellite in the next communicationIt is clear that self and fragment->Fragments of the same index, and fragments->Polymerization is carried out.
Similarly, the low-orbit satellite receives the segment of the last communication low-orbit satellite transmissionAt the same time as the last communication low-orbit satellite transmission segment +.>Index of (2) to self and fragment->Fragments of the same index, and fragments->Polymerization is carried out.
According to the federal learning method based on the low-orbit satellite constellation, the index of the segment is transmitted simultaneously when the segment is transmitted, so that the low-orbit satellite in the next communication can definitely aggregate the received segment with the same index.
In some embodiments, the federal learning method based on low-orbit satellite constellation provided in the embodiments of the present invention further includes:
fragments are processedThe corresponding satellite index is transmitted to the low-orbit satellite in the next communication for the low-orbit satellite in the next communication, in the and segment +.>Pairs of identical index fragmentsThe corresponding satellite index, contained in the fragment->In the case of the corresponding satellite index, the corresponding fragment +.>Substitution of the same-indexed fragment with fragment->
Receiving segments of low-orbit satellite transmissions in a previous communicationA corresponding satellite index;
at and segmentThe satellite index corresponding to the segment with the same index is contained in the segment +.>In the case of the corresponding satellite index, the corresponding fragment +.>Substitution of the same-indexed fragment with fragment->
In particular, due to the presence of transmitted fragmentsComprises and fragments->In the case of segments of the same index, therefore, for the purpose of determining the transmitted segment +.>Whether or not to contain and/or block->Fragments of the same index, fragments +.>The corresponding satellite index is transmitted to the low-orbit satellite in the next communication.
At and segmentThe satellite index corresponding to the segment with the same index is contained in the segment +.>In the case of the corresponding satellite index, the fragment +. >Comprises and fragments->Fragments of the same index, to avoid duplication, will itself be identical to fragmentsSubstitution of the same-indexed fragment with fragment->
Similarly, the low-orbit satellite receives a segment of the low-orbit satellite transmission in the last communicationCorresponding satellite index, in the and segment->The satellite index corresponding to the segment with the same index is contained in the segment +.>In the case of the corresponding satellite index, the fragment +.>Comprises and fragments->Fragments of the same index will be identical to fragment +.>Substitution of the same-indexed fragment with fragment->
According to the federal learning method based on the low-orbit satellite constellation, the satellite indexes corresponding to the fragments are transmitted simultaneously when the fragments are transmitted, so that whether the fragments definitely received by the low-orbit satellite in the next communication contain the fragments with the same indexes or not is ensured, and further, in the case that the received fragments contain the fragments with the same indexes, the fragments with the same indexes are replaced with the received fragments in order to avoid repetition.
In order to more clearly understand the process of implementing in-orbit aggregation by multiple rounds of transmission, a detailed description will be given below with specific examples.
Exemplary, low-orbit satellite S 1 、S 2 And S is 3 The annular communication sequence between the two is low-orbit satellite S 1 Transmitted to low-orbit satellite S 2 Low orbit satellite S 2 Transmitted to low-orbit satellite S 3 And low orbit satellite S 3 Transmitted to low-orbit satellite S 1 . Low orbit satellite S 1 、S 2 And S is 3 All the model updating parameters are split into 3 fragments, and the low orbit satellite S 1 Respectively of 3 fragments ofAnd->Low orbit satellite S 2 Is +.>And->Low orbit satellite S 3 Is +.>And->
In the first round of transmission, a low-orbit satellite S 1 Fragments are processedFragment->Fragment index 1 and fragment->The corresponding satellite index {1} is transmitted to the low-orbit satellite S 2 Low orbit satellite S 2 The received fragment->And fragment->Polymerization is carried out to obtainAs a new fragment 3.
Low orbit satellite S 2 Fragments are processedFragment->Fragment index 2 and fragment->The corresponding satellite index {2} is transmitted to the low-orbit satellite S 3 Low orbit satellite S 3 The received fragment->And fragment->Polymerizing to obtain->As a new fragment 2.
Low orbit satellite S 3 Fragments are processedFragment->Fragment index 1 and fragment->Corresponding satellite index {3} is transmitted to low-orbit satellite S 1 Low orbit satellite S 1 The received fragment->And fragment->Polymerizing to obtain->As a new fragment 1.
After the first round of transmission, the low-orbit satellite S 1 Respectively of 3 fragments ofAnd->Low orbit satellite S 2 Is +.>And->Low orbit satellite S 3 Is +.>And
in the second round of transmission, the low-orbit satellite S 1 Fragments are processedFragment->Fragment index 1 and fragment->Corresponding satellite indexes {1,3} are transmitted to the low-orbit satellite S 2 Low orbit satellite S 2 Fragments to be receivedAnd fragment->Polymerizing to obtain->As a new fragment 1.
Low orbit satellite S 1 Fragments are processedFragment->Fragment index 3 and fragment->The corresponding satellite index {1,2} is transmitted to the low-orbit satellite S 3 Low orbit satellite S 3 The received fragment->And fragment->Polymerizing to obtain->As a new fragment 3.
Low orbit satellite S 3 Fragments are processedFragment->Fragment index 2 and fragment->Corresponding satellite indexes {2,3} are transmitted to the low-orbit satellite S 1 Low orbit satellite S 1 The received fragment->And fragment->Polymerization is carried outObtain->As a new fragment 2.
After the second round of transmission, the low-orbit satellite S 1 Respectively of 3 fragments of And->Low orbit satellite S 2 Is +.>And->Low orbit satellite S 3 Is +.>And->
In the third round of transmission, the low-orbit satellite S 1 Fragments are processedFragment->Fragment index 2 and fragment->The corresponding satellite index {1,2,3} is transmitted to the low-orbit satellite S 2 Low orbit satellite S 2 The received fragment->And fragment->Aggregation is performed, and the received fragments are used for avoiding repetitionDirect substitution of the original fragment->I.e. fragment->As a new fragment 2.
Low orbit satellite S 2 Fragments are processedFragment->Fragment index 1 of (2) and fragmentsThe corresponding satellite index {1,2,3} is transmitted to the low-orbit satellite S 3 Low orbit satellite S 3 Fragments to be receivedAnd fragment->Polymerization is carried out, the received fragments are left in order to avoid repetition>Direct substitution of the original fragment->I.e.)>As a new fragment 1.
Low orbit satellite S 3 Fragments are processedFragment->Fragment index 3 and fragment->The corresponding satellite index {1,2,3} is transmitted to the low-orbit satellite S 1 Low orbit satellite S 1 Fragments to be receivedAnd fragment->Polymerization is carried out, the received fragments are left in order to avoid repetition>Direct substitution of the original fragment->I.e.)>As a new fragment 3.
After the third round of transmission, low-orbit satellite S 1 Respectively of 3 fragments of Andlow orbit satellite S 2 Is +.>Andlow orbit satellite S 3 Is +.>And->
In the third round of transmission, the low-orbit satellite S 1 Fragments are processedFragment->Fragment index 3 and fragment->The corresponding satellite index {1,2,3} is transmitted to the low-orbit satellite S 2 Low orbit satellite S 2 The received fragment- >And fragment->Aggregation is performed, and the received fragments are used for avoiding repetitionDirect substitution of the original fragment->I.e. fragment->As a new fragment 3.
Low orbit satellite S 2 Fragments are processedFragment->Fragment index 2 of (2) and fragmentsThe corresponding satellite index {1,2,3} is transmitted to the low-orbit satellite S 3 Low orbit satellite S 3 Fragments to be receivedAnd fragment->Polymerization is carried out, the received fragments are left in order to avoid repetition>Direct substitution of the original fragment->I.e.)>As a new fragment 2.
Low orbit satellite S 3 Fragments are processedFragment->Fragment index 3 of (2) and fragmentsThe corresponding satellite index {1,2,3} is transmitted to the low-orbit satellite S 1 Low orbit satelliteS 1 Fragments to be receivedAnd fragment->Polymerization is carried out, the received fragments are left in order to avoid repetition>Direct substitution of the original fragment->I.e.)>As a new fragment 1.
After the fourth round of transmission, the low-orbit satellite S 1 Respectively of 3 fragments of And->Low orbit satellite S 2 Is +.>Andlow orbit satellite S 3 Is +.>Andto this end, low-orbit satellite S 1 、S 2 And S is 3 All the model updating parameters are held, and the in-orbit aggregation is completed.
In some embodiments, the federal learning method based on low-orbit satellite constellation provided in the embodiments of the present invention further includes:
In the case that it is determined that the track is unable to realize annular communication and that the track itself is able to communicate with the ground station, the local model update is performed and then the own model update parameters are transmitted to the ground station.
In particular, in the case where the low-orbit satellite is in an orbit in which annular communication is not possible and itself can communicate with the ground station, i.e., S nei =0 and S g In the case of =1, since ring communication cannot be achieved, there is also no in-orbit aggregation process, and the federal learning process includes: local model updating and global aggregation.
The low orbit satellite directly transmits own model updating parameters to the ground station after carrying out local model updating, and the ground station immediately executes global aggregation after receiving the model updating parameters. The process of local model update and global aggregation is as described above and will not be described in detail herein.
According to the federal learning method based on the low-orbit satellite constellation, provided by the embodiment of the invention, under the condition that the orbit can not realize annular communication and can communicate with the ground station, after the local model is updated, the update parameters of the model are transmitted to the ground station, so that the time efficiency of the federal learning algorithm in the low-orbit satellite constellation is further improved.
In some embodiments, the federal learning method based on low-orbit satellite constellation provided in the embodiments of the present invention further includes:
local model updates are made in the event that it is determined that the in-orbit is unable to effect annular communication and that all of the low-orbit satellites in the in-orbit are unable to communicate with the ground station.
In particular, in the case where the low-orbit satellites are in orbits in which annular communication cannot be achieved, and all of the low-orbit satellites in orbit cannot communicate with the ground station, i.e., S nei =0 and S g In the case of =0, since loop-through cannot be achievedNor is it able to communicate with the ground station, and therefore, there is also no in-orbit and global aggregation process, and the federal learning process includes: and updating the local model.
The low-orbit satellite only performs local model updating, and the process of local model updating is as described above and will not be described here again. Waiting for the subsequent state change, and executing the intra-track aggregation under the condition that the track can realize annular communication; in the case of communication with the ground station, global aggregation is performed again.
According to the federal learning method based on the low-orbit satellite constellation, local model updating is performed under the condition that the orbit cannot realize annular communication and all low-orbit satellites in the orbit cannot communicate with a ground station, so that the state change can execute subsequent operations.
The low-orbit satellite constellation-based federal learning device provided by the invention is described below, and the low-orbit satellite constellation-based federal learning device described below and the low-orbit satellite constellation-based federal learning method described above can be correspondingly referred to each other.
Fig. 3 is a schematic structural diagram of a federal learning device based on low-orbit satellite constellation, and referring to fig. 3, the invention provides a federal learning device based on low-orbit satellite constellation, which comprises: a first processing module 310.
The first processing module 310 is configured to, when it is determined that the orbit can implement annular communication and there is a converging satellite in the orbit, transmit its own model update parameter to the converging satellite after performing local model update, so that the converging satellite transmits the model update parameters of all low-orbit satellites in the orbit to a ground station;
wherein the converged satellite is a low-orbit satellite in which the communicable time with the ground station is longer than the sum of the converged time and the transmission time in the orbit; the convergence time is the time required for converging the model update parameters of all the low-orbit satellites in the orbit; the transmission time is the time required for transmitting the aggregated model update parameters to the ground station.
In some embodiments, the apparatus includes a second processing module;
and the second processing module is used for transmitting the model updating parameters to the low-orbit satellite in the next communication according to the communication direction of the annular communication after carrying out local model updating under the condition that the orbit can realize annular communication and all the low-orbit satellites in the orbit can not communicate with the ground station, and receiving the model updating parameters transmitted by the low-orbit satellite in the last communication until each low-orbit satellite in the orbit holds the model updating parameters of all the low-orbit satellites.
In some embodiments, the second processing module comprises: a split sub-module and a first transmission sub-module; wherein:
the splitting module is used for splitting the update parameters of the model into p fragments;
the first transmission sub-module is used for transmitting the fragments in the ith round of transmissionTransmitting to said low-orbit satellite in the next communication and receiving the fragment of the low-orbit satellite transmission in the last communication>
Wherein the saidUpdating parameters for the model of the xth segment of its own p segments, where x is p-k- (i [ mod ] ]p) +2, said k being the ordering of itself in said track; said->Updating parameters for a model of a y-th segment of the p segments of the last-communicated low-orbit satellite, wherein y is x+1[ mod ]]p, the k-1 is the order of the low-orbit satellite in the last communication in the orbit.
In some embodiments, the second processing module further comprises: the system comprises a second transmission sub-module, a first receiving sub-module and an aggregation sub-module; wherein:
the second transmission sub-module is used for transmitting the fragmentsIs transmitted to the low-orbit satellite in the next communication for the low-orbit satellite in the next communication to be associated with the fragment +.>Fragments of the same index, identical to said fragments->Polymerizing;
the first receiving sub-module is used for receiving the segment of the low-orbit satellite transmission in the last communicationIndex of (2);
the aggregation sub-module is used for connecting with the fragmentsFragments of the same index, identical to said fragments->Polymerization is carried out.
In some embodiments, the second processing module further comprises: the system comprises a third transmission sub-module, a second receiving sub-module and a replacing sub-module; wherein:
the third transmission sub-module is used for transmitting the fragments Transmitting the corresponding satellite index to the low-orbit satellite in the next communication for the low-orbit satellite in the next communication to communicate with the segment/>Satellite indexes corresponding to the same-index fragments are contained in the fragments +.>In the case of the corresponding satellite index, will be +_with the fragment>Substitution of the same indexed fragment with said fragment +.>
The second receiving sub-module is used for receiving the segment of the low-orbit satellite transmission in the last communicationA corresponding satellite index;
the replacing sub-module is used for replacing the segmentSatellite indexes corresponding to the same-index fragments are contained in the fragments +.>In the case of the corresponding satellite index, will be +_with the fragment>Substitution of the same indexed fragment with said fragment +.>
In some embodiments, the apparatus further comprises: a third processing module;
and the third processing module is used for transmitting the update parameters of the own model to the ground station after the local model update is carried out under the condition that the track can not realize annular communication and the own model can be communicated with the ground station.
In some embodiments, the apparatus further comprises: a fourth processing module;
The fourth processing module is configured to perform a local model update if it is determined that the in-orbit cannot achieve annular communication, and that all low-orbit satellites in the in-orbit cannot communicate with the ground station.
It should be noted that, the federal learning device based on the low-orbit satellite constellation provided by the invention can realize all the method steps realized by the method embodiment and can achieve the same technical effects, and the parts and beneficial effects same as those of the method embodiment in the embodiment are not specifically repeated here.
Fig. 4 is a schematic structural diagram of an electronic device according to the present invention, as shown in fig. 4, the electronic device may include: processor 410, communication interface (Communications Interface) 420, memory 430 and communication bus 440, wherein processor 410, communication interface 420 and memory 430 communicate with each other via communication bus 440. The processor 410 may invoke logic instructions in the memory 430 to perform a federal learning method based on a low-orbit satellite constellation, the method comprising: under the condition that the orbit can realize annular communication and a converging satellite exists in the orbit, after local model updating is carried out, transmitting model updating parameters of the converging satellite to the converging satellite so that the converging satellite can transmit model updating parameters of all low-orbit satellites in the orbit to a ground station;
Wherein the converged satellite is a low-orbit satellite in which the communicable time with the ground station is longer than the sum of the converged time and the transmission time in the orbit; the convergence time is the time required for converging the model update parameters of all the low-orbit satellites in the orbit; the transmission time is the time required for transmitting the aggregated model update parameters to the ground station.
Further, the logic instructions in the memory 430 described above may be implemented in the form of software functional units and may be stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
In another aspect, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being storable on a non-transitory computer readable storage medium, the computer program, when executed by a processor, being capable of performing the low-orbit satellite constellation-based federal learning method provided by the methods described above, the method comprising: under the condition that the orbit can realize annular communication and a converging satellite exists in the orbit, after local model updating is carried out, transmitting model updating parameters of the converging satellite to the converging satellite so that the converging satellite can transmit model updating parameters of all low-orbit satellites in the orbit to a ground station;
wherein the converged satellite is a low-orbit satellite in which the communicable time with the ground station is longer than the sum of the converged time and the transmission time in the orbit; the convergence time is the time required for converging the model update parameters of all the low-orbit satellites in the orbit; the transmission time is the time required for transmitting the aggregated model update parameters to the ground station.
In yet another aspect, the present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, is implemented to perform the low-orbit satellite constellation-based federal learning method provided by the above methods, the method comprising: under the condition that the orbit can realize annular communication and a converging satellite exists in the orbit, after local model updating is carried out, transmitting model updating parameters of the converging satellite to the converging satellite so that the converging satellite can transmit model updating parameters of all low-orbit satellites in the orbit to a ground station;
Wherein the converged satellite is a low-orbit satellite in which the communicable time with the ground station is longer than the sum of the converged time and the transmission time in the orbit; the convergence time is the time required for converging the model update parameters of all the low-orbit satellites in the orbit; the transmission time is the time required for transmitting the aggregated model update parameters to the ground station.
The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.
From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (10)
1. A federal learning method based on low-orbit satellite constellation, comprising:
under the condition that the orbit can realize annular communication and a converging satellite exists in the orbit, after local model updating is carried out, transmitting model updating parameters of the converging satellite to the converging satellite so that the converging satellite can transmit model updating parameters of all low-orbit satellites in the orbit to a ground station;
wherein the converged satellite is a low-orbit satellite in which the communicable time with the ground station is longer than the sum of the converged time and the transmission time in the orbit; the convergence time is the time required for converging the model update parameters of all the low-orbit satellites in the orbit; the transmission time is the time required for transmitting the aggregated model update parameters to the ground station.
2. The low-orbit satellite constellation-based federal learning method according to claim 1, further comprising:
and under the condition that the orbit can realize annular communication and all the low-orbit satellites in the orbit can not communicate with the ground station, after local model updating is carried out, transmitting own model updating parameters to the low-orbit satellites in the next communication according to the communication direction of the annular communication, and receiving the model updating parameters transmitted by the low-orbit satellites in the last communication until each low-orbit satellite in the orbit holds the model updating parameters of all the low-orbit satellites.
3. The federal learning method based on low-orbit satellite constellation according to claim 2, wherein the transmitting the model update parameters to the low-orbit satellite in the next communication and receiving the model update parameters transmitted from the low-orbit satellite in the previous communication according to the communication direction of the ring communication comprises:
splitting the update parameters of the self model into p segments;
in the ith round of transmission, fragments are transmittedTransmitting to said low-orbit satellite in the next communication and receiving the fragment of the low-orbit satellite transmission in the last communication >
Wherein the saidUpdating parameters for the model of the xth segment of its own p segments, where x is p-k- (i [ mod ]]p) +2, said k being the ordering of itself in said track; said->Updating parameters for a model of a y-th segment of the p segments of the last-communicated low-orbit satellite, wherein y is x+1[ mod ]]p, the k-1 is the order of the low-orbit satellite in the last communication in the orbit.
4. A method of federal learning based on low-earth satellite constellations according to claim 3, further comprising:
the fragments are subjected toThe index of the next position is transmitted to the next positionA low-orbit satellite for communication, so that the low-orbit satellite for the next communication will be in communication with the fragment +.>Fragments of the same index, identical to said fragments->Polymerizing;
receiving the segment of the low-orbit satellite transmission in the last communicationIndex of (2);
will be associated with the fragmentFragments of the same index, identical to said fragments->Polymerization is carried out.
5. The low-orbit satellite constellation based federal learning method according to claim 4, further comprising:
the fragments are subjected toTransmitting the corresponding satellite index to the low-orbit satellite in the next communication for the low-orbit satellite in the next communication, wherein the low-orbit satellite is in the same way as the fragment ∈ - >Satellite indexes corresponding to the same-index fragments are contained in the fragments +.>In the corresponding satellite index, will be as describedFragment->Substitution of the same indexed fragment with said fragment +.>
Receiving the segment of the low-orbit satellite transmission in the last communicationA corresponding satellite index;
at the same time as the fragmentSatellite indexes corresponding to the same-index fragments are contained in the fragments +.>In the case of the corresponding satellite index, will be +_with the fragment>Substitution of the same indexed fragment with said fragment +.>
6. The low-orbit satellite constellation-based federal learning method according to claim 1, further comprising:
and under the condition that the track can not realize annular communication and can communicate with the ground station, transmitting the model updating parameters to the ground station after carrying out local model updating.
7. The low-orbit satellite constellation-based federal learning method according to claim 1, further comprising:
a local model update is performed upon determining that the in-orbit cannot achieve annular communication and that all low-orbit satellites in the in-orbit cannot communicate with the ground station.
8. A federal learning device based on low orbit satellite constellation, comprising:
the first processing module is used for transmitting the model updating parameters of the first processing module to the converging satellites after carrying out local model updating under the condition that the orbit can realize annular communication and the converging satellites exist in the orbit, so that the converging satellites can transmit the model updating parameters of all low-orbit satellites in the orbit to the ground station;
wherein the converged satellite is a low-orbit satellite in which the communicable time with the ground station is longer than the sum of the converged time and the transmission time in the orbit; the convergence time is the time required for converging the model update parameters of all the low-orbit satellites in the orbit; the transmission time is the time required for transmitting the aggregated model update parameters to the ground station.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the low orbit satellite constellation based federal learning method according to any one of claims 1 to 7 when the computer program is executed by the processor.
10. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the federal learning method based on low-earth satellite constellations according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311533269.1A CN117455014A (en) | 2023-11-16 | 2023-11-16 | Federal learning method and device based on low orbit satellite constellation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311533269.1A CN117455014A (en) | 2023-11-16 | 2023-11-16 | Federal learning method and device based on low orbit satellite constellation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117455014A true CN117455014A (en) | 2024-01-26 |
Family
ID=89583546
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311533269.1A Pending CN117455014A (en) | 2023-11-16 | 2023-11-16 | Federal learning method and device based on low orbit satellite constellation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117455014A (en) |
-
2023
- 2023-11-16 CN CN202311533269.1A patent/CN117455014A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110443375B (en) | Method and device for federated learning | |
| Han et al. | Adaptive gradient sparsification for efficient federated learning: An online learning approach | |
| Razmi et al. | On-board federated learning for dense LEO constellations | |
| CN113010305B (en) | Federal learning system deployed in edge computing network and learning method thereof | |
| KR102247322B1 (en) | Method for Operating Machine Learning Based Federated Distillation, Web Server and Terminal | |
| EP3979717A1 (en) | Method for association of user equipment in a cellular network according to a transferable association policy | |
| CN117455014A (en) | Federal learning method and device based on low orbit satellite constellation | |
| Cao et al. | Multi-tier collaborative deep reinforcement learning for non-terrestrial network empowered vehicular connections | |
| CN104780567A (en) | Satellite network flow control method | |
| CN116133082A (en) | Multi-hop clustering method for improving topology duration of aviation ad hoc network | |
| CN111835640B (en) | Shortest time delay routing method based on continuous time aggregation graph | |
| CN115587266A (en) | Air-space-ground integrated internet intelligent edge caching method | |
| CN115173926A (en) | Communication method and communication system of satellite-ground converged relay network based on auction mechanism | |
| US20230246887A1 (en) | Training in Communication Systems | |
| CN114268348A (en) | Honeycomb-free large-scale MIMO power distribution method based on deep reinforcement learning | |
| Elmahallawy et al. | Communication-Efficient Federated Learning for LEO Satellite Networks Integrated with HAPs using Hybrid NOMA-OFDM | |
| WO2024037321A1 (en) | Ai model training method and apparatus, and device and readable storage medium | |
| Zhou et al. | DRL-Based Workload Allocation for Distributed Coded Machine Learning | |
| CN116980881B (en) | Multi-unmanned aerial vehicle collaboration data distribution method, system, electronic equipment and medium | |
| WO2023283785A1 (en) | Method for processing signal, and receiver | |
| WO2022233061A1 (en) | Signal processing method, communication device, and communication system | |
| CN116264704B (en) | Low-power-consumption wide area network sense fusion method based on channel sensing and reinforcement learning | |
| Bos et al. | Avoiding normalization uncertainties in deep learning architectures for end-to-end communication | |
| CN116366152A (en) | Relay auxiliary multichannel FSO network dynamic resource allocation method based on MP-DQN | |
| CN115600681A (en) | Intelligent model training 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 |