A star-ground connection density-aware federated learning method based on asynchronous dynamic aggregation

Abstract

The application discloses a star-ground connection density-aware federated learning method and system based on asynchronous dynamic aggregation, and a storage medium. The method is applied to a ground station and comprises the following steps: S1, starting the i-th round of updating, and sending a current global model to the i-th group of satellites in a satellite constellation; S2, judging whether a storage queue meets a first preset condition or whether a current updating duration meets a second preset condition; if yes, executing S3; if no, executing S2; S3, extracting all local data packets in the storage queue and updating the current global model; S4, when judging that i does not reach a preset updating round number, regarding all satellites corresponding to all local data packets as the i+1-th group of satellites, and then returning to S1 after assigning i+1 to i. The asynchronous dynamic aggregation introduced in the application can solve the problems of long satellite waiting time caused by excessively sparse connection density of satellites and ground stations and the waste of satellite computing resources caused by excessively dense connection of satellites and ground stations.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ground station applications, in particular to a star-ground connection density-aware federated learning method and system based on asynchronous dynamic aggregation and a storage medium. BACKGROUND

[0002] With the development of space technology, the phenomenon of using satellites for earth observation is becoming more and more common. In-orbit satellites are gradually evolving into a new computing paradigm for future computer systems because their calculations are not constrained by complex changes in the earth's surface and natural disasters. With the efforts of various disciplines, satellite technology has made significant progress and reduced the price of satellite deployment, among which low earth orbit satellites have attracted widespread attention due to their short return period and low cost.

[0003] However, due to technical limitations, traditional satellite systems cannot process data in orbit and can only transmit data containing a large amount of useless information back to the ground station, and then the ground station transmits the data to the cloud computing platform to train the machine learning model, at which time the download of all generated raw image data on the satellite brings huge pressure to the downstream bandwidth. Based on this, some scholars propose to apply federated learning to satellite constellations, but because the satellite and the ground station are in relative motion, the connection between them is dynamically changing, which makes the federated learning have the following problems: first, when the connection density between the satellite and the ground station is too sparse, it will cause other satellites to spend a lot of time waiting; second, when the connection between the satellite and the ground station is too dense, it will cause a large number of satellites to be in an idle state, causing waste of computing resources. SUMMARY

[0004] The present application provides a star-ground connection density-aware federated learning method and system based on asynchronous dynamic aggregation and a storage medium to solve one or more technical problems existing in the prior art and at least provide a beneficial choice or create conditions.

[0005] In a first aspect, a star-ground connection density-aware federated learning method based on asynchronous dynamic aggregation is provided, applied to a ground station, the method comprising:

[0006] Step 100, start the i-th round of update, and send the current global model to the i-th group of satellites in the satellite constellation;

[0007] Step 200, real-time judge whether the storage queue satisfies the first preset condition or the current update duration satisfies the second preset condition, the storage queue is used to real-time sequence cache the local data packet fed back by one of the satellites in the satellite constellation received; if yes, execute step 300; if no, execute step 200;

[0008] Step 300, extracting all local data packets in the storage queue and updating the current global model;

[0009] Step 400, when determining that i does not reach the preset update round, taking all satellites corresponding to the all local data packets as the i+1 group of satellites, and then assigning i+1 to i and returning to step 100.

[0010] Further, the first preset condition is that the number of the local data packets currently cached in the storage queue reaches a preset buffer capacity.

[0011] Further, the second preset condition is that the current update duration reaches an integer multiple of a preset time period.

[0012] Further, in the step 200, the satellite is one of the i group of satellites, or is a satellite that receives the current global model sent by the ground station in the previous round of update.

[0013] Further, in the step 200, the generation process of the local data packet fed back by the satellite includes:

[0014] When the satellite receives the current global model sent by the ground station, the current update round of the ground station is recorded;

[0015] The satellite initializes its local model by using the received current global model, and trains the local model by using its local data set to obtain an optimal local model;

[0016] The optimal local model, the local data set size, and the update round are data encapsulated to form a local data packet.

[0017] Further, the step 300 includes:

[0018] According to the update round recorded by each local data packet, the weight value of the optimal local model carried by each local data packet is calculated, and each optimal local model and its weight value are combined to update the current global model;

[0019] According to the local data set size recorded by each local data packet, the loss function of the optimal local model carried by each local data packet is weighted and summed to obtain the loss function of the updated current global model.

[0020] Further, the combination of each optimal local model and its weight value to update the current global model is corresponding to the update formula:

[0021]

[0022] Among them, w i Let w be the current global model obtained after completing the i-th round of updates. i-1 Let α be the current global model obtained after completing the (i-1)th round of updates, α be the hybrid hyperparameter, and B be all the local data packets. k w represents the weight value of the optimal local model carried in the k-th local data packet. k For the optimal local model carried by the k-th local data packet, τ k This represents the update round recorded in the k-th local data packet.

[0023] Furthermore, the loss function of the optimal local model carried in each local data packet is weighted and summed, and the corresponding summation formula is as follows:

[0024]

[0025] Where, f(w) i ) represents the current global model w i The loss function, m k Let f be the size of the local dataset recorded in the k-th local data packet, m be the sum of the sizes of all local datasets recorded in all the local data packets, and f be the size of the local dataset recorded in the k-th local data packet. k (w k ) is the optimal local model w k The loss function, D k x is the local dataset associated with the satellite that is part of the k-th local data packet. k For local dataset D k One of the data points, l(w) k ,x k ) is the optimal local model w k For data x k Training loss.

[0026] Secondly, a density-aware federated learning system for satellite-to-ground connectivity based on asynchronous dynamic aggregation is provided. The system includes a ground station and a satellite constellation, wherein the satellite constellation contains a number of satellites.

[0027] The ground station is used to initiate the i-th update round by sending the current global model to the i-th group of satellites in the satellite constellation; it determines in real time whether the storage queue meets the first preset condition or whether the current update duration meets the second preset condition. The storage queue is used to cache the local data packets received from one of the satellites in the satellite constellation in real time. If not, it continues to perform the current real-time judgment operation. If yes, it extracts all the local data packets in the storage queue and updates the current global model; it determines whether i has reached the preset update round number. If not, it takes all the satellites corresponding to all the local data packets as the (i+1)-th group of satellites, assigns i+1 to i, and returns to start a new round of update. If yes, it ends the update process.

[0028] Each satellite, upon receiving the current global model sent by the ground station, records the current update round of the ground station; initializes its own local model using the received current global model; trains the local model using its own local dataset to obtain the optimal local model; encapsulates the optimal local model, the size of the local dataset, and the update round into a local data packet and feeds it back to the ground station.

[0029] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the space-to-ground connectivity density-aware federated learning method based on asynchronous dynamic aggregation as described in the first aspect.

[0030] This invention offers at least the following advantages: By introducing an asynchronous dynamic aggregation strategy into federated learning applied to satellite constellations, which integrates asynchronous periodic aggregation and asynchronous buffered aggregation strategies, when the connection density between satellites and ground stations is too sparse, the ground station can complete an aggregation update operation even when the storage queue is not full but a preset time period has been reached, thereby shortening the update waiting time for other satellites. Conversely, when the connection density between satellites and ground stations is too dense, the ground station can complete an aggregation update operation even before the preset time period has been reached, even when the storage queue is full, thus fully utilizing the computing resources of a large number of idle satellites. Compared to traditional asynchronous or synchronous aggregation strategies, the asynchronous dynamic aggregation strategy proposed in this invention can significantly save satellite waiting time. Furthermore, when the ground station performs a global model update, it fully utilizes the update rounds previously participated in by all associated satellites, avoiding unnecessary impacts on the optimization of the global model and effectively ensuring the model optimization results. Attached Figure Description

[0031] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0032] Figure 1 This is a flowchart illustrating a density-aware federated learning method for satellite-to-ground connectivity based on asynchronous dynamic aggregation, as described in an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram showing a comparative analysis of simulation test results of federated learning methods under four different aggregation strategies in embodiments of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0035] It should be noted that although functional modules are divided in the system diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first," "second," "third," etc., used in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed and are inherent to these processes, methods, products, or apparatuses.

[0036] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a density-aware federated learning method for satellite-to-ground connectivity based on asynchronous dynamic aggregation, provided by an embodiment of the present invention. Applied to ground stations, the method includes the following steps:

[0037] S110. Start the i-th round of updates and send the current global model to the i-th group of satellites in the satellite constellation;

[0038] S120. In real time, determine whether the storage queue meets the first preset condition or whether the current update duration meets the second preset condition. The storage queue is used to cache the local data packets fed back by one of the satellites in the satellite constellation in real time in sequence. If yes (i.e., at least one of the first preset condition and the second preset condition is met), then continue to execute the following step S130. If no (i.e., neither the first preset condition nor the second preset condition is met), then execute the step S120 again.

[0039] S130. Extract all local data packets in the storage queue and update the current global model, thereby ending the i-th round of updates;

[0040] S140. Determine whether i < N is true, where N is the preset update round number; if yes, proceed to step S150; if no, proceed to step S160.

[0041] S150. Take all the satellites corresponding to all the local data packets as the (i+1)th group of satellites, and then assign i+1 to i. At this time, return to execute the above step S110.

[0042] S160. End the update process for the global model.

[0043] It should be noted that the above step S110 is executed starting from i=1 (i is a positive integer), and the storage queue is empty when the first round of updates is started. The ground station will randomly send the initial global model to some satellites that have already established communication connections. These satellites are defined as the first group of satellites.

[0044] In this embodiment of the invention, when any satellite is visible to the ground station, the satellite can establish a communication connection with the ground station, specifically through the following communication model:

[0045]

[0046] Where, r k (t) represents the satellite's trajectory in the geocentric inertial coordinate system, r g (t) represents the trajectory of the ground station in the geocentric inertial coordinate system, where t is a continuous clock time, and ∠[r g (t),r k (t)-r g [t] refers to the satellite trajectory r k (t) and ground station trajectory r g The trajectory difference formed by (t) and the ground station trajectory r g The angle between (t) can also be expressed as the angle between the satellite and the ground station and denoted as α. k,g (t), α minThis represents the minimum elevation angle of the ground station; that is, the above communication model actually represents the minimum elevation angle α of the satellite. min When the satellite is visible from the ground station, the link between the satellite and the ground station is feasible.

[0047] It should be noted that the geocentric inertial coordinate system refers to a coordinate system that remains stationary or in uniform linear motion (without acceleration) in space. The origin is the center of the Earth, the Z-axis coincides with the Earth's rotation axis, the X-axis lies in the equatorial plane and points to the prime meridian, and the Y-axis lies in the equatorial plane and points to the meridian at 90° east longitude.

[0048] In this embodiment of the invention, the first preset condition mentioned in step S120 actually refers to the number of local data packets currently cached in the storage queue reaching a preset buffer capacity, which is the maximum number of data packets that the storage queue is allowed to cache, as set in advance by the technician.

[0049] In this embodiment of the invention, the second preset condition mentioned in step S120 above actually refers to the current update duration being an integer multiple of the preset time period. The preset time period is the time interval for each update specified by the ground station and set in advance by the technician. The current update duration refers to the specific duration elapsed from the start of the first round of updates by the ground station until the current moment (in the i-th round of updates).

[0050] In this embodiment of the invention, since the time taken for each satellite in the satellite constellation to complete local training is different, during the i-th round of updates, when the local training speed of any one or more satellites in the i-th group is relatively fast, the ground station may receive local data packets fed back by any one or more satellites in the i-th group and cache them into the storage queue in a timely manner; or, during the i-th round of updates, when the local training speed of the satellites of the current global model sent by the ground station in any one or several previous rounds of updates is relatively slow, the ground station may receive local data packets fed back by these satellites and cache them into the storage queue in a timely manner.

[0051] In this embodiment of the invention, the corresponding generation process for the local data packets fed back by the satellite mentioned in step S120 above includes the following steps:

[0052] (1) After the satellite establishes a communication connection with the ground station for the first time, it receives the current global model sent by the ground station and obtains the current update round of the ground station.

[0053] (2) The satellite will initialize its local model based on the received current global model, then call its local dataset and combine it with the SGD (Stochastic Gradient Descent) algorithm to train the initialized local model, thereby obtaining the optimal local model; the corresponding training process is as follows:

[0054]

[0055] Among them, w j+1 w represents the local model obtained by the satellite after performing the (j+1)th iteration of training. j Let η be the local model obtained by the satellite after performing the j-th iteration of training, and η be the learning rate. Let Y represent the gradient operator. j f(w) is a batch dataset randomly drawn by the satellite from its own locally stored dataset during the j-th iteration of training. j ,Y j ) for the local model w j The loss function, B1 is the batch dataset Y j The amount of data contained, y j For batch dataset Y j One of the data points, l(w) j ,y j ) for the local model w j For data y j Training loss;

[0056] (3) The satellite encapsulates the update round, the size of the local dataset, and the optimal local model to generate a local data packet, and then sends the local data packet after establishing a communication connection with the ground station for the next time.

[0057] In this embodiment of the invention, due to the potential asynchronous training between the ground station and the satellite, the ground station may have already undergone multiple rounds of updates by the time the satellite completes its local training. This results in a time difference between the optimal local model obtained after the satellite completes its local training and the current global model to be updated by the ground station. Furthermore, stale local gradients can severely reduce the utility of the current global model. To address this technical problem, a staleness function is introduced in step S130 to facilitate model updates. The specific implementation process includes the following steps:

[0058] S131. For each local data packet extracted from the storage queue, calculate the weight value of the optimal local model it carries using the staleness function, based on its recorded update round:

[0059]

[0060] In the formula, α k τ represents the weight value of the optimal local model carried in the k-th local data packet, α is the hybrid hyperparameter with a value range of α∈(0,1), i is the current update round of the ground station, and τ k For the update round recorded in the k-th local data packet, s(i-τ) k ) represents the obsolescence function used to solve for the model obsolescence reflected by the k-th local data packet, thereby adjusting the value of the hybrid hyperparameter α; when i = τ k This indicates that during the i-th round of updates at the ground station, the satellite associated with the k-th local data packet just received the current global model sent by the ground station and quickly completed the local training and local data packet feedback tasks. When i ≠ τ k (Only i>τ is possible) k When the satellite associated with the k-th local data packet completes its local training and local data packet feedback tasks, the ground station has already completed the i-τ process. k Rotational updates;

[0061] S132. Based on the weight values ​​of each optimal local model calculated in step S131 above, the current global model is updated using each optimal local model, specifically as follows:

[0062] w i =(1-α)w i - 1 + k∈B α k w k ;

[0063] In the formula, w i Let w be the current global model obtained by the ground station after performing the i-th round of updates. i-1 Let w be the current global model obtained by the ground station after performing the (i-1)th round of updates. k B represents the optimal local model carried by the k-th local data packet, and B refers to all local data packets extracted from the storage queue.

[0064] S133. Combining the local dataset size recorded in each local data packet with the loss function of the optimal local model, the updated loss function of the current global model is obtained by weighted summation:

[0065]

[0066] In the formula, f(w) i) represents the current global model w mentioned above. i The loss function, m k Let f be the size of the local dataset recorded in the k-th local data packet, m be the sum of the sizes of all local datasets recorded in all the local data packets, and f be the size of the local dataset recorded in the k-th local data packet. k (w k The optimal local model w is described above. k The loss function is calculated directly on the satellite corresponding to the k-th local data packet, D. k Let x be the local dataset provided by the satellite corresponding to the k-th local data packet. k For local dataset D k One of the data points, l(w) k ,x k The optimal local model w is described above. k For data x k The resulting training loss.

[0067] In existing technologies, strategies for implementing density-aware federated learning for satellite-to-ground connectivity include at least the traditional asynchronous aggregation strategy, the asynchronous periodic aggregation strategy, and the asynchronous buffered aggregation strategy, which are described in detail below:

[0068] Traditional asynchronous aggregation strategy: When multiple satellites in the satellite constellation feed back their best local models trained by themselves to the ground station, the ground station will use the single best local model received in the order of reception to complete the update operation of the current global model.

[0069] Asynchronous periodic aggregation strategy: By pre-setting an update time period for the ground station, when one or more satellites in the satellite constellation feed back their best local model trained within a single update time period to the ground station, the ground station will use all the best local models received within the single update time period to complete the update operation of the current global model at the end of the single update time period.

[0070] Asynchronous buffer aggregation strategy: By pre-setting a storage queue for the ground station, when multiple satellites in the satellite constellation feed back their best local models trained by themselves to the ground station, the ground station will use all the best local models cached in the storage queue to complete the update operation of the current global model when it detects that the storage queue used for real-time caching of the best local models is full.

[0071] The asynchronous periodic aggregation strategy is designed to address the issue that when the connection density between satellites and ground stations is too sparse, a set update time period is used to periodically trigger the ground stations to aggregate all the best local models they have received in order to update the current global model. Compared with traditional asynchronous aggregation strategies, this saves update time and alleviates the challenge of long update waiting times caused by sparse satellite connections.

[0072] The asynchronous buffer aggregation strategy is designed to address the issue of excessively dense connections between satellites and ground stations. By setting up a storage queue with fixed storage space, it quantitatively triggers the ground station to aggregate all received optimal local models to update the current global model. Compared to traditional asynchronous aggregation strategies, this approach better addresses the heterogeneity of satellite connections over time, thereby alleviating the challenge of wasting idle satellite computing resources caused by dense satellite connections.

[0073] To verify the relative superiority of the asynchronous dynamic aggregation strategy proposed in this invention, this embodiment uses the existing COTE simulator (a simulator developed by foreign scholars that can physically model orbital mechanics and Earth's rotation to track the positions of ground stations and satellites over time) to simulate a satellite network consisting of a satellite constellation and 10 ground stations. The satellite constellation consists of 100 satellites. The space portion of the satellite network includes a polar (97.3°) orbit containing the satellite constellation, and the ground portion includes a first circle of ground stations located at 87° North latitude and a second circle of ground stations located at 87° South latitude. Each circle of ground stations has 5 ground stations arranged in a vertically uniform distribution, and any ground station can establish a communication connection with each satellite in the satellite constellation.

[0074] Based on this, this embodiment of the invention calls upon existing Non-IID (Not identically and independently distributed) datasets and divides them into training and testing sets. Taking the implementation of density-aware federated learning of satellite-ground connectivity between a ground station and the satellite constellation according to the four different aggregation strategies mentioned above as an example, the training set is randomly allocated to each satellite in the satellite constellation to facilitate local training. Then, the testing set is used to simulate and test the federated learning results under various aggregation strategies. See details below. Figure 2 As shown, the asynchronous dynamic aggregation strategy proposed in this invention can achieve better federated learning results compared with the other three types of aggregation strategies in the prior art, and the required training time is significantly shortened.

[0075] In this embodiment of the invention, an asynchronous dynamic aggregation strategy is introduced into federated learning applied to satellite constellations. This strategy combines asynchronous periodic aggregation and asynchronous buffered aggregation. When the connection density between satellites and ground stations is too sparse, the ground station can complete an aggregation update operation before the storage queue is full, thus shortening the update waiting time of other satellites. Conversely, when the connection density between satellites and ground stations is too dense, the ground station can complete an aggregation update operation before the preset time period is reached, even if the storage queue is full, thus fully utilizing the computing resources of a large number of idle satellites. Compared to traditional asynchronous or synchronous aggregation strategies, the asynchronous dynamic aggregation strategy proposed in this invention can significantly save satellite waiting time. Furthermore, when the ground station performs a global model update, it fully utilizes the update rounds previously participated in by all associated satellites, avoiding unnecessary impacts on the optimization of the global model and effectively ensuring the model optimization results.

[0076] In addition, this invention also provides a space-to-ground connection density-aware federated learning system based on asynchronous dynamic aggregation. The system includes a ground station and a satellite constellation consisting of several satellites, wherein each satellite can establish a communication connection with the ground station independently. The specific functions of the two types of equipment are as follows:

[0077] The ground station is used to send its stored current global model to the i-th group of satellites in the satellite constellation when the i-th update round is started, where i is a positive integer; to determine in real time whether the storage queue meets a first preset condition or whether the current update duration meets a second preset condition, wherein the storage queue can cache local data packets fed back by one of the satellites in the satellite constellation in real time in sequence; if neither the first preset condition nor the second preset condition is met, the station continues to return to perform this real-time judgment operation; if at least one of the first preset condition and the second preset condition is met, all local data packets are extracted from the storage queue to complete the update operation of the current global model; to determine whether i < N is true, where N is the preset update round number; if true, all satellites corresponding to all local data packets are defined as the (i+1)-th group of satellites, and then i+1 is assigned to i before returning to start a new round of update operation; if false, the update process of the global model ends.

[0078] Wherein, the first preset condition actually refers to the number of local data packets currently cached in the storage queue reaching the preset buffer capacity, the preset buffer capacity being the maximum number of data packets allowed to be cached by the storage queue as set in advance by the technicians; the second preset condition actually refers to the current update duration being an integer multiple of the preset time period, the preset time period being the time interval between each update specified by the ground station as set in advance by the technicians, and the current update duration referring to the specific duration elapsed from the start of the first round of updates by the ground station until the current moment (in the i-th round of updates).

[0079] After establishing an initial communication connection with the ground station, any satellite receives the current global model sent by the ground station and obtains the current update round of the ground station. Based on the received current global model, it initializes its own local model, then uses its local dataset and SGD (Stochastic Gradient Descent) algorithm to train the initialized local model, thereby obtaining the optimal local model. The update round, local dataset size, and the optimal local model are then encapsulated to generate a local data packet, which is then sent after the next communication connection with the ground station is established.

[0080] In this embodiment of the invention, when any satellite is visible to the ground station, the satellite can establish a communication connection with the ground station, specifically through the following communication model:

[0081]

[0082] Where, r k (t) represents the satellite's trajectory in the geocentric inertial coordinate system, r g (t) represents the trajectory of the ground station in the geocentric inertial coordinate system, where t is a continuous clock time, and ∠[r g (t),r k (t)-r g [t] refers to the satellite trajectory r k (t) and ground station trajectory r g The trajectory difference formed by (t) and the ground station trajectory r g The angle between (t) can also be expressed as the angle between the satellite and the ground station and denoted as α. k,g (t), α min This represents the minimum elevation angle of the ground station; that is, the above communication model actually represents the minimum elevation angle α of the satellite. min When the satellite is visible from the ground station, the link between the satellite and the ground station is feasible.

[0083] It should be noted that the geocentric inertial coordinate system refers to a coordinate system that remains stationary or in uniform linear motion (without acceleration) in space. The origin is the center of the Earth, the Z-axis coincides with the Earth's rotation axis, the X-axis lies in the equatorial plane and points to the prime meridian, and the Y-axis lies in the equatorial plane and points to the meridian at 90° east longitude.

[0084] The content of the above method embodiments is applicable to this system embodiment. The functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are the same as those in the above method embodiments. Therefore, they will not be repeated here.

[0085] Furthermore, embodiments of the present invention also provide a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the asynchronous dynamic aggregation-based satellite-to-ground connectivity density-aware federated learning method described in the above embodiments. The computer-readable storage medium includes, but is not limited to, any type of disk (including floppy disks, hard disks, optical disks, CD-ROMs, and magneto-optical disks), ROM (Read-Only Memory), RAM (Random Access Memory), EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, magnetic cards, or optical cards. In other words, the storage device includes any medium on which a device (e.g., a computer, mobile phone, etc.) stores or transmits information in a readable form, and can be a read-only memory, a disk, or an optical disk, etc.

[0086] Although the description of this application has been quite detailed and particularly focused on several of the described embodiments, it is not intended to limit itself to any of these details or embodiments or any particular embodiment. Rather, it should be considered as effectively covering the intended scope of this application by referring to the appended claims and taking into account the prior art, which provides for a broad possible interpretation of these claims. Furthermore, the foregoing description of this application with respect to embodiments foreseeable by the inventors is intended to provide a useful description, and non-substantial modifications to this application that have not yet been foreseen may still represent equivalent modifications.

Claims

1. A density-aware federated learning method for satellite-to-ground connectivity based on asynchronous dynamic aggregation, characterized in that, Applied to ground stations, the method includes: Step 100: Initiate the i-th round of updates and send the current global model to the i-th group of satellites in the satellite constellation; Step 200: In real time, determine whether the storage queue meets the first preset condition or whether the current update duration meets the second preset condition. The storage queue is used to cache local data packets received from one of the satellites in the satellite constellation in real time and in sequence. The first preset condition is that the number of local data packets currently cached in the storage queue reaches a preset buffer capacity. The second preset condition is that the current update duration reaches an integer multiple of a preset time period. If yes, proceed to step 300; if no, proceed to step 200. Step 300: Extract all local data packets from the storage queue and update the current global model, including: Based on the update round recorded in each local data packet, the weight value of the optimal local model carried in each local data packet is calculated. Then, the current global model is updated by combining each optimal local model and its weight value. The corresponding update formula is as follows: In the formula, This is the current global model obtained after completing the i-th round of updates. This is the current global model obtained after completing the (i-1)th round of updates. For hybrid hyperparameters, For all the local data packets, This represents the weight value of the optimal local model carried in the k-th local data packet. The optimal local model carried by the k-th local data packet. This refers to the update round recorded in the k-th local data packet; Based on the size of the local dataset recorded in each local data packet, the loss function of the optimal local model carried in each local data packet is weighted and summed to obtain the updated loss function of the current global model. The corresponding summation formula is as follows: In the formula, For the current global model loss function, Let k be the size of the local dataset recorded in the k-th local data packet. The sum of the sizes of all local datasets recorded in all the local data packets. For the optimal local model loss function, This refers to the local dataset associated with the k-th local data packet from the satellite. For local datasets One of the data, For the optimal local model Data Training loss; Step 400: When it is determined that i has not reached the preset number of update rounds, all satellites corresponding to all local data packets are taken as the (i+1)th group of satellites, and then i+1 is assigned to i before returning to step 100.

2. The satellite-to-ground connectivity density-aware federated learning method based on asynchronous dynamic aggregation according to claim 1, characterized in that, In step 200, the satellite is either one of the satellites in the i-th group, or a satellite that received the current global model sent by the ground station in a previous round of updates.

3. The satellite-to-ground connectivity density-aware federated learning method based on asynchronous dynamic aggregation according to claim 1, characterized in that, In step 200, the process of generating the local data packets fed back by the satellite includes: When the satellite receives the current global model sent by the ground station, it records the current update round of the ground station; The satellite initializes its own local model using the received current global model, and then trains the local model using its own local dataset to obtain the optimal local model. The optimal local model, the local dataset size, and the update round are encapsulated to form a local data package.

4. A density-aware federated learning system for satellite-to-ground connectivity based on asynchronous dynamic aggregation, characterized in that, The system includes a ground station and a satellite constellation, the satellite constellation containing a number of satellites; The ground station is used to initiate the i-th round of updates and send the current global model to the i-th group of satellites in the satellite constellation. The system continuously checks whether the storage queue meets a first preset condition or whether the current update duration meets a second preset condition. The storage queue is used to cache local data packets received from a satellite in the satellite constellation in real time. The first preset condition is that the number of local data packets currently cached in the storage queue reaches a preset buffer capacity. The second preset condition is that the current update duration reaches an integer multiple of a preset time period. If not, the current real-time judgment operation continues. If yes, all local data packets in the storage queue are extracted and the current global model is updated. The system then checks whether i has reached a preset update round. If not, all satellites corresponding to all local data packets are taken as the (i+1)th group of satellites, and i+1 is assigned to i before returning to start a new round of updates. If yes, the update process ends. Each satellite, upon receiving the current global model sent by the ground station, records the current update round of the ground station; initializes its own local model using the received current global model; trains the local model using its own local dataset to obtain the optimal local model; encapsulates the optimal local model, the size of the local dataset, and the update round into a local data packet and feeds it back to the ground station. The step of extracting all local data packets in the storage queue and updating the current global model includes: Based on the update round recorded in each local data packet, the weight value of the optimal local model carried in each local data packet is calculated. Then, the current global model is updated by combining each optimal local model and its weight value. The corresponding update formula is as follows: In the formula, This is the current global model obtained after completing the i-th round of updates. This is the current global model obtained after completing the (i-1)th round of updates. For hybrid hyperparameters, For all the local data packets, This represents the weight value of the optimal local model carried in the k-th local data packet. The optimal local model carried by the k-th local data packet. This refers to the update round recorded in the k-th local data packet; Based on the size of the local dataset recorded in each local data packet, the loss function of the optimal local model carried in each local data packet is weighted and summed to obtain the updated loss function of the current global model. The corresponding summation formula is as follows: In the formula, For the current global model loss function, Let k be the size of the local dataset recorded in the k-th local data packet. The sum of the sizes of all local datasets recorded in all the local data packets. For the optimal local model loss function, This refers to the local dataset associated with the k-th local data packet from the satellite. For local datasets One of the data, For the optimal local model Data Training loss.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the space-to-ground connectivity density-aware federated learning method based on asynchronous dynamic aggregation as described in any one of claims 1 to 3.

Images