Communication efficient federated object detection method and system for low earth orbit satellite constellation

Abstract

The application discloses a communication efficient federated target detection method and system for a low-orbit satellite cluster, relates to the technical field of distributed computing, and comprises the following steps: initializing a target detection model on a master satellite and distributing the target detection model to client satellites; performing local training on the target detection model on the client satellites, obtaining updated gradients, and compressing the updated gradients according to a preset compression mode selection strategy to obtain compressed gradients; selecting client satellites participating in federated aggregation in the current round on the master satellite based on a preset selection strategy, and sending an aggregation instruction to the selected client satellites; decompressing the received compressed gradients on the master satellite, updating the model based on the decompressed gradients and a synchronous-asynchronous hybrid aggregation mechanism, and obtaining an updated target detection model; and calculating the difference between the model before and after the update and sending the difference to all client satellites, so that the client satellites update the model. The application can realize rapid deployment and update of the target detection model on the satellite.

Description

Technical Field

[0001] This invention relates to the field of distributed computing technology, and in particular to a method and system for efficient federated target detection for low-Earth orbit satellite constellations. Background Technology

[0002] With the rapid development and large-scale deployment of satellite technology, low-Earth orbit (LEO) satellite constellations have become crucial infrastructure for Earth observation, environmental monitoring, and emergency response. LEO satellites can continuously acquire remote sensing data with wide coverage and high spatiotemporal resolution, providing a massive data source for target detection missions. Traditional satellite data processing employs a centralized architecture of "satellite acquisition → satellite-to-ground transmission → ground processing," where satellites transmit raw remote sensing data back to the ground via satellite-to-ground links, and then high-performance computing clusters on the ground perform centralized processing. However, this satellite data processing method suffers from the following problems when dealing with massive amounts of data: Limited bandwidth resources in satellite-to-ground links and long transmission times for massive amounts of remote sensing data make it difficult to meet the real-time and response speed requirements of scenarios such as disaster monitoring and target tracking. Centralized ground processing requires data to be transmitted from satellites, and the long transmission cycle leads to poor system timeliness. Cross-domain and cross-network segment transmission of remote sensing image data may lead to data leakage issues.

[0003] To address the aforementioned issues, federated learning, as a distributed machine learning approach, has been introduced into the field of satellite computing. The core idea of ​​federated learning is "data doesn't move, model moves," meaning each node trains its model locally using its private data, only transmitting model update information (such as gradients or parameters) to the central node. This allows for the collaborative training of a global machine learning model without sharing the original data. However, directly applying federated learning to the highly dynamic and resource-constrained environment of low-Earth orbit satellite constellations presents the following problems: Inter-satellite networks are highly dynamic and unstable: there are only periodic, brief visual windows between satellites and ground stations, and inter-satellite links also exhibit strong intermittency and instability due to the high-speed movement of satellites. Traditional federated learning usually relies on synchronous communication protocols, requiring all participating nodes to complete local training and upload updates within a specified time. In satellite networks, this can easily lead to training stagnation or a sharp drop in efficiency due to node disconnection.

[0004] Limited onboard resources: Satellites have more limited resources such as computing, storage, and energy than ground equipment. Running large-scale target detection models and conducting multiple rounds of local training will place a heavy burden on onboard resources.

[0005] The problem of highly heterogeneous data distribution: Due to differences in orbital position, sensor type, imaging time and climate conditions, the data acquired by different satellites vary in terms of regional coverage, spectral characteristics and target distribution, which leads to slow convergence or decreased generalization performance of federated learning models.

[0006] Communication overhead issues: With a large constellation and a large number of model parameters, even transmitting only model update information will still consume considerable inter-satellite link resources, which may affect the normal operation of other missions. Summary of the Invention

[0007] To address some or all of the technical problems existing in the prior art, this invention provides a communication-efficient federated target detection method and system for low-Earth orbit satellite constellations. This method can significantly reduce communication overhead and improve training efficiency while protecting data privacy, and adapt to the high dynamics and resource-constrained characteristics of satellite networks. This enables rapid deployment and updating of target detection models on satellites, thereby enhancing the on-orbit collaborative intelligent processing capabilities of low-Earth orbit satellite constellations.

[0008] The technical solution of the present invention is as follows: Firstly, a highly efficient federated target detection method for low-Earth orbit satellite constellations is provided, the method comprising: Load and initialize the target detection model on the master satellite, and distribute the target detection model to the client satellites; The target detection model is trained locally on the client satellite using locally stored remote sensing data to obtain the corresponding update gradient; The inter-satellite link quality is calculated on the client satellite. Based on the inter-satellite link quality and the preset compression method, a strategy is selected to compress the update gradient, the compressed gradient is obtained, and the compression error is calculated and stored. Based on the status information of the client satellites, the main satellite selects the client satellites to participate in the federated aggregation in this round according to the preset selection strategy, and sends the aggregation command to the selected client satellites so that the selected client satellites send the compression gradient and the corresponding metadata to the main satellite. The received compressed gradients are decompressed on the main satellite, and the target detection model is updated based on the decompressed gradients and the synchronous-asynchronous hybrid aggregation mechanism to obtain the updated target detection model. The synchronous-asynchronous hybrid aggregation mechanism includes: synchronously aggregating gradients received within a specified time and asynchronously fusing gradients that arrive after the timeout but match based on time decay. Calculate the model difference between the target detection model before and after the update, and send the model difference to all client satellites or selected client satellites participating in the next round of federated aggregation, so that the client satellites can update the target detection model according to the model difference and carry out the next round of target detection model training and update until the training requirements are met.

[0009] Furthermore, in some implementations, the quality of the inter-satellite link is calculated in the following manner: ; in, Indicates the quality of inter-satellite links. This represents the normalized value of the inter-satellite link bandwidth availability between the client satellite and the master satellite. This represents the normalized value of the round-trip time delay of the inter-satellite link between the client satellite and the master satellite. This indicates the packet loss rate of the inter-satellite link between the client satellite and the master satellite. The weight representing the availability of inter-satellite link bandwidth. The weights representing the round-trip delay of inter-satellite links. The weights represent the packet loss rate of inter-satellite links.

[0010] Furthermore, in some embodiments, the preset compression method selection strategy includes: When the quality of the inter-satellite link is less than the first quality threshold, the gradient is compressed using the Top-K gradient sparsity method, and the K value in the Top-K gradient sparsity method is set to the first value. When the quality of the inter-satellite link is not less than the first quality threshold and not greater than the second quality threshold, the gradient is compressed using the Top-K gradient sparsification method, and the K value in the Top-K gradient sparsification method is set to the second value. When the quality of the inter-satellite link is greater than the second quality threshold, the gradient is compressed using the low-rank matrix factorization method, and the rank value in the low-rank matrix factorization method is set to the third value.

[0011] Furthermore, in some embodiments, the compression error is calculated in the following manner: ; in, Indicates compression error. This indicates that the gradient is being updated. This represents the compression gradient.

[0012] Furthermore, in some implementations, the preset selection strategy includes: From the client satellites whose status update interval does not exceed the preset time, select the client satellites whose remaining energy is not lower than the preset energy threshold, whose computing load is not higher than the preset load threshold, and whose available memory is not lower than the preset memory threshold; Based on the selected client satellites, and considering their remaining energy, computational load, and remote sensing data quality, a corresponding importance score is calculated. The selected client satellites are sorted from highest to lowest importance score, and the top P client satellites are selected as the client satellites participating in this round of federated aggregation, where P is a preset positive integer.

[0013] Furthermore, in some implementations, the corresponding importance score is calculated using the following formula: ; in, Indicates the importance score, This indicates the remaining energy of the client satellite. This indicates the computing load of the client satellite. This indicates the quality of the client's satellite remote sensing data. This represents the weight corresponding to the remaining energy. This indicates the weight corresponding to the calculated load. This represents the weight corresponding to the quality of remote sensing data.

[0014] Furthermore, in some implementations, updating the target detection model based on the decompressed gradient and the synchronous-asynchronous hybrid aggregation mechanism includes: Synchronously aggregate the gradients received within a specified time to obtain the synchronously aggregated gradients; The target detection model is updated based on the synchronously aggregated gradient; Calculate the corresponding decaying learning rate based on the delay time of the matched gradient that arrives after the timeout. The object detection model is updated again based on the gradients that arrive after the timeout and match, and their corresponding decaying learning rates.

[0015] Furthermore, in some implementations, the gradients received within a specified time are synchronously aggregated using the following formula to obtain the synchronously aggregated gradient: ; Based on the synchronously aggregated gradient, the target detection model is updated using the following formula; ; Based on the delay time corresponding to the matched gradient that arrives after a timeout, the corresponding decaying learning rate is calculated using the following formula: ; Based on the timed-out and matched gradients and their corresponding decaying learning rates, the object detection model is updated again using the following formula: ; in, This indicates synchronous gradient aggregation. Indicates the first The decompressed gradient express The corresponding amount of training data, Indicates the number of client satellites used for synchronization aggregation. This represents the updated object detection model parameters. This represents the object detection model parameters before the update. Indicates the learning rate. This represents the decaying learning rate. Represents an exponential function. Indicates the attenuation coefficient. Indicates the delay time. This represents the updated object detection model parameters. This represents the gradient that has reached its timeout and is matched after decompression.

[0016] Furthermore, in some embodiments, the method further includes: when performing the next round of training and updating of the target detection model, after obtaining the corresponding update gradient, acquiring the locally stored compression error, updating the momentum accumulation error according to the compression error, and using the momentum accumulation error to correct the update gradient obtained by local training. The momentum accumulation error is updated using the following formula based on the compression error: ; in, This represents the cumulative momentum error. This represents the cumulative momentum error from the previous round. Indicates the momentum coefficient. Indicates compression error; The update gradient obtained from local training is corrected using the following formula: ; in, This indicates that the gradient is being updated. This represents the corrected updated gradient.

[0017] Secondly, a high-efficiency federated target detection system for low-Earth orbit satellite constellations is also provided. The system includes: a model initialization unit, a client selection and scheduling unit, and a first model update processing unit deployed on the master satellite; and a local training unit, a dynamic gradient compression processing unit, a local caching unit, and a second model update processing unit deployed on the client satellite. The model initialization unit is used to load and initialize the target detection model, and distribute the target detection model to the client satellite; The local training unit is used to train the target detection model locally using locally stored remote sensing data to obtain the corresponding update gradient. The dynamic gradient compression processing unit is used to calculate the inter-satellite link quality, select a strategy based on the inter-satellite link quality and a preset compression method, select a compression method to compress the updated gradient to obtain the compressed gradient, calculate the compression error, and send the compressed gradient and corresponding metadata to the main satellite according to the aggregation instruction. The local cache unit is used to store compression errors; The client selection and scheduling unit is used to select client satellites participating in the federated aggregation in this round based on the status information of the client satellites and a preset selection strategy, and send aggregation instructions to the selected client satellites; The first model update processing unit is used to decompress the received compressed gradient, update the target detection model based on the decompressed gradient and the synchronous-asynchronous hybrid aggregation mechanism, and obtain the updated target detection model. The synchronous-asynchronous hybrid aggregation mechanism includes: synchronously aggregating gradients received within a specified time, asynchronously fusing gradients that arrive after timeout and match based on time decay; and calculating the model difference between the target detection model before and after the update, and sending the model difference to all client satellites or selected client satellites participating in the next round of federated aggregation. The second model update processing unit is used to update the target detection model based on the received model difference.

[0018] The main advantages of the technical solution of this invention are as follows: The efficient federated target detection method and system for low-Earth orbit (LEO) satellite constellations of this invention, through dynamic gradient compression and error compensation, resource-aware client satellite selection, and the synergy of synchronous-asynchronous hybrid aggregation mechanisms, can systematically solve the problems of high dynamics and instability of inter-satellite networks, limited onboard resources, highly heterogeneous data distribution, and communication overhead faced by federated learning in LEO satellite constellation environments, while protecting the data security of each satellite. It can significantly reduce inter-satellite communication overhead, ensure model convergence accuracy and target detection performance, greatly improve the efficiency and robustness of federated learning in dynamic and unstable networks, and enhance the on-orbit collaborative intelligent processing capabilities of LEO satellite constellations. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and constitute a part of this invention, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 A flowchart of an efficient federated target detection method for low-Earth orbit satellite constellations provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] The technical solutions provided by the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] refer to Figure 1 In a first aspect, embodiments of the present invention provide a method for efficient federated target detection in low-Earth orbit satellite constellations, the method comprising the following steps: Step 1: Load and initialize the target detection model on the master satellite, and distribute the target detection model to the client satellites; Step 2: Train the target detection model locally on the client satellite using locally stored remote sensing data to obtain the corresponding update gradient; Step 3: Based on the real-time sensing calculation of the inter-satellite link quality on the client satellite, select a strategy according to the inter-satellite link quality and the preset compression method, compress the update gradient according to the selected compression method, obtain the compressed gradient, and calculate and store the compression error. Step 4: On the main satellite, based on the status information of the client satellites, select the client satellites to participate in the federated aggregation in this round according to the preset selection strategy, and send the aggregation command to the selected client satellites so that the selected client satellites send the compression gradient and the corresponding metadata to the main satellite. Step 5: Decompress the received compressed gradient on the main satellite, update the target detection model based on the decompressed gradient and the synchronous-asynchronous hybrid aggregation mechanism, and obtain the updated target detection model. The synchronous-asynchronous hybrid aggregation mechanism includes: synchronously aggregating the gradients received within a specified time, and asynchronously fusing the gradients that arrive after the timeout and match based on time decay. Step 6: Calculate the model difference between the target detection model before and after the update, and send the model difference to all client satellites or selected client satellites participating in the next round of federated aggregation, so that the client satellites can update the target detection model according to the model difference and carry out the next round of target detection model training and update until the training requirements are met.

[0023] In this embodiment of the invention, the master satellite is any designated satellite or mission execution satellite in the low-Earth orbit satellite constellation, and the client satellite is any other satellite in the low-Earth orbit satellite constellation other than the master satellite.

[0024] In this embodiment of the invention, the target detection model loaded on the main satellite is specifically set according to actual needs, for example, the YOLO series target detection model can be used.

[0025] In this embodiment of the invention, during local training, remote sensing data is input into the target detection model to obtain the detection results output by the model. Based on these results, the gradient of the loss function value relative to the target detection model parameters is calculated as the corresponding update gradient. The loss function used during training is specifically set according to actual needs. For example, when the target detection model uses the YOLO series target detection model, the loss function is the existing loss function used for training YOLO series target detection models.

[0026] In this embodiment of the invention, the client satellite periodically sends its own status information to the master satellite, and the master satellite updates the status information of the client satellite in real time based on the received information.

[0027] In this embodiment of the invention, the metadata corresponding to the compression gradient includes: the amount of training data corresponding to the update gradient and the compression method. The amount of training data represents the amount of remote sensing data used to obtain the update gradient.

[0028] In this embodiment of the invention, the client satellite packages the compression gradient and the corresponding metadata into a data packet and sends it to the main satellite.

[0029] In this embodiment of the invention, the specified time is set according to the actual situation.

[0030] In this embodiment of the invention, if the updated version of the target detection model corresponding to the compressed gradient is the same as the updated version of the target detection model on the current main satellite, it indicates a match; if the updated version of the target detection model corresponding to the compressed gradient is different from the updated version of the target detection model on the current main satellite, it indicates a mismatch. Gradients that arrive after a timeout and are mismatched are discarded.

[0031] In this embodiment of the invention, the training requirements are set according to actual needs, for example, the number of training updates for the object detection model is set to a certain number.

[0032] In this embodiment of the invention, steps 2 to 6 are repeated multiple times to train and update the target detection model in multiple rounds.

[0033] The efficient federated target detection method for low-Earth orbit (LEO) satellite constellations provided in this invention, through dynamic gradient compression and error compensation, resource-aware client satellite selection, and the synergy of synchronous-asynchronous hybrid aggregation mechanisms, can systematically solve the problems of high dynamics and instability of inter-satellite networks, limited onboard resources, highly heterogeneous data distribution, and communication overhead faced by federated learning in LEO satellite constellation environments, while protecting the data security of each satellite. It can significantly reduce inter-satellite communication overhead, ensure model convergence accuracy and target detection performance, greatly improve the efficiency and robustness of federated learning in dynamic and unstable networks, and enhance the on-orbit collaborative intelligent processing capabilities of LEO satellite constellations.

[0034] Furthermore, in one embodiment of the present invention, in step 3, the inter-satellite link quality is calculated in the following manner: ; in, Indicates the quality of inter-satellite links. This represents the normalized value of the inter-satellite link bandwidth availability between the client satellite and the master satellite. This represents the normalized value of the round-trip time delay of the inter-satellite link between the client satellite and the master satellite. This indicates the packet loss rate of the inter-satellite link between the client satellite and the master satellite. The weight representing the availability of inter-satellite link bandwidth. The weights representing the round-trip delay of inter-satellite links. The weights represent the packet loss rate of inter-satellite links.

[0035] Among them, weight , , The specific values ​​are set according to actual needs, and .

[0036] The normalized value can be calculated using the minimum-maximum scaling method or the z-score normalization method.

[0037] The client satellite obtains the inter-satellite link bandwidth availability, inter-satellite link round-trip latency, and inter-satellite link packet loss rate in real time through the underlying communication protocol stack.

[0038] In this embodiment of the invention, by integrating the three dynamic indicators of inter-satellite link bandwidth, latency, and packet loss rate into a single inter-satellite link quality, a decision-making basis can be provided for the selection of subsequent gradient compression methods. This enables an effective balance between optimizing the utilization of communication resources and ensuring the reliability of model update transmission in an environment where the inter-satellite link status changes in real time.

[0039] Furthermore, in one embodiment of the present invention, in step 3, the preset compression method selection strategy includes: When the quality of the inter-satellite link is less than the first quality threshold, the gradient is compressed using the Top-K gradient sparsity method, and the K value in the Top-K gradient sparsity method is set to the first value. When the quality of the inter-satellite link is not less than the first quality threshold and not greater than the second quality threshold, the gradient is compressed using the Top-K gradient sparsification method, and the K value in the Top-K gradient sparsification method is set to the second value. When the quality of the inter-satellite link is greater than the second quality threshold, the gradient is compressed using the PowerSGD low-rank matrix factorization method, and the rank value in the low-rank matrix factorization method is set to the third value.

[0040] In this embodiment of the invention, the principle of the Top-K gradient sparsification method is to sort the absolute values ​​of all elements in the gradient from largest to smallest, and only retain the values ​​of the first K% of elements.

[0041] In this embodiment of the invention, the principle of the low-rank matrix factorization method is to approximately decompose the gradient into the product of two smaller matrices. For example, if the magnitude of the gradient is m×n, then the gradient is approximately decomposed into the product of an m×rank matrix and a rank×n matrix, where rank is a predetermined rank value.

[0042] In this embodiment of the invention, the first quality threshold, the second quality threshold, the first value, the second value, and the third value are specifically set according to actual needs, for example, they are set to 0.3, 0.7, 0.5, 2, and 8 respectively.

[0043] Furthermore, in this embodiment of the invention, the compression error is calculated in the following manner: ; in, Indicates compression error. This indicates that the gradient is being updated. This represents the compression gradient.

[0044] In this embodiment of the invention, by dividing the quality of the inter-satellite link into three intervals and matching them with three different compression methods, the compression intensity and the link carrying capacity are adapted. Under extremely poor link conditions, connectivity is prioritized, while high-fidelity transmission is pursued under good link conditions. Thus, under any link conditions, the communication load can be reduced to the maximum extent, while providing accurate gradient information for subsequent model updates, ensuring the optimal balance between model update efficiency and accuracy.

[0045] Furthermore, in one embodiment of the present invention, in step 4, the preset selection strategy includes: From the client satellites whose status update interval does not exceed the preset time, select the client satellites whose remaining energy is not lower than the preset energy threshold, whose computing load is not higher than the preset load threshold, and whose available memory is not lower than the preset memory threshold; Based on the selected client satellites, and considering their remaining energy, computational load, and remote sensing data quality, a corresponding importance score is calculated. The selected client satellites are sorted from highest to lowest importance score, and the top P client satellites are selected as the client satellites participating in this round of federated aggregation, where P is a preset positive integer.

[0046] In this embodiment of the invention, the status update time interval represents the time interval at which the client satellite sends status information to the master satellite.

[0047] In this embodiment of the invention, the preset time is set according to actual needs.

[0048] In this embodiment of the invention, the preset energy threshold, preset load threshold, and preset memory threshold are set according to actual needs.

[0049] Furthermore, in this embodiment of the invention, the corresponding importance score is calculated using the following formula based on the remaining energy of the client satellite, computing load, and remote sensing data quality: ; in, Indicates the importance score, This indicates the remaining energy of the client satellite. This indicates the computing load of the client satellite. This indicates the quality of the client's satellite remote sensing data. This represents the weight corresponding to the remaining energy. This indicates the weight corresponding to the calculated load. This represents the weight corresponding to the quality of remote sensing data.

[0050] In this embodiment of the invention, remaining energy, computational load, and remote sensing data quality are expressed as percentage values. Remote sensing data quality is determined based on the correlation between remote sensing data and the currently executed task, with weights... , , The specific values ​​are set according to actual needs, and .

[0051] In this embodiment of the invention, the specific value of P is set according to actual needs.

[0052] In this embodiment of the invention, by comprehensively considering the energy sufficiency, computing idleness and data value of the client satellite for quantitative scoring and selection, the training failure or performance drag caused by resource-depleted satellites is effectively avoided. High-value data is prioritized for aggregation, thereby maximizing the efficiency and stability of each round of federated learning under the constraints of satellite resources, laying a key foundation for the rapid and high-quality convergence of the target detection model.

[0053] Furthermore, in one embodiment of the present invention, in step 5, the target detection model is updated based on the decompressed gradient and the synchronous-asynchronous hybrid aggregation mechanism, including: Synchronously aggregate the gradients received within a specified time to obtain the synchronously aggregated gradients; The target detection model is updated based on the synchronously aggregated gradient; Calculate the corresponding decaying learning rate based on the delay time of the matched gradient that arrives after the timeout. The object detection model is updated again based on the gradients that arrive after the timeout and match, and their corresponding decaying learning rates.

[0054] In this embodiment of the invention, the gradients received within a specified time are synchronously aggregated using the following formula to obtain the synchronously aggregated gradient: ; in, This indicates synchronous gradient aggregation. Indicates the first The decompressed gradient express The corresponding amount of training data, This indicates the number of client satellites used for synchronization aggregation.

[0055] In this embodiment of the invention, the target detection model is updated using the following formula based on the synchronously aggregated gradient; ; in, This represents the updated object detection model parameters. This represents the object detection model parameters before the update. This represents the learning rate, which is set according to actual needs.

[0056] In this embodiment of the invention, the corresponding decay learning rate is calculated using the following formula based on the delay time corresponding to the gradient that arrives after a timeout and is matched: ; in, This represents the decaying learning rate. Indicates the learning rate. Represents an exponential function. Indicates the attenuation coefficient. This represents the delay time, and the decay coefficient is a positive number. The specific value is set according to actual needs and is used to determine the rate at which the learning rate decays with the delay time.

[0057] In this embodiment of the invention, the target detection model is updated again based on the gradients that arrive after the timeout and are matched, and their corresponding decaying learning rates, using the following formula: ; in, This represents the updated object detection model parameters. This represents the gradient that has reached its timeout and is matched after decompression.

[0058] In this embodiment of the invention, by using the above-mentioned synchronous-asynchronous hybrid aggregation mechanism to update the target detection model, it is possible to fully utilize the computing resources of lagging client satellites in the highly dynamic and intermittent network of low-Earth orbit satellite constellations, while effectively avoiding the interference of outdated gradients on the model optimization direction, thereby achieving a synergistic improvement in the overall throughput, robustness and model accuracy of federated learning.

[0059] Furthermore, in one embodiment of the present invention, in step 6, the model difference is calculated in the following manner: ; in, Indicates the model difference. Indicates the first The target detection model is trained and updated in rounds. Indicates the first The target detection model is trained and updated in rounds.

[0060] In this embodiment of the invention, the client satellite updates the target detection model based on the model difference in the following manner: ; in, Indicates the first During each training and update cycle, the target detection model on the client satellite... Indicates the first The target detection model on the client satellite is updated during each training cycle.

[0061] Furthermore, in one embodiment of the present invention, step 6 may further include: The model difference is losslessly or lossily compressed, and the compressed model difference is sent to all client satellites or selected client satellites participating in the next round of federated aggregation, so that the client satellites can decompress the received model difference and update the target detection model based on the decompressed model difference.

[0062] In this embodiment of the invention, by compressing the model difference, the amount of inter-satellite communication data can be further reduced.

[0063] Furthermore, in one embodiment of the present invention, the method further includes: when performing the next round of training and updating of the target detection model, after obtaining the corresponding update gradient, acquiring the locally stored compression error, updating the momentum accumulation error according to the compression error, and using the momentum accumulation error to correct the update gradient obtained by local training.

[0064] In this embodiment of the invention, the momentum accumulation error is updated using the following formula based on the compression error: ; in, This represents the cumulative momentum error. This represents the cumulative momentum error from the previous round. Indicates the momentum coefficient. This indicates compression error.

[0065] The initial value of the momentum accumulation error is set to 0, meaning that the momentum accumulation error is 0 during the first round of training updates.

[0066] Among them, momentum coefficient Configure according to actual needs. .

[0067] In this embodiment of the invention, the update gradient obtained from local training is corrected using the following formula: ; in, This indicates that the gradient is being updated. This represents the corrected updated gradient.

[0068] In this embodiment of the invention, by continuously and smoothly feeding back historical compression errors to subsequent training gradients in the form of momentum, the information loss caused by high-proportion gradient compression is effectively offset, thereby maintaining model convergence accuracy while ensuring communication efficiency.

[0069] Secondly, embodiments of the present invention also provide a communication-efficient federated target detection system for low-Earth orbit satellite constellations. The system includes: a model initialization unit, a client selection and scheduling unit, and a first model update processing unit deployed on the master satellite; and a local training unit, a dynamic gradient compression processing unit, a local caching unit, and a second model update processing unit deployed on the client satellite. The model initialization unit is used to load and initialize the target detection model and distribute the target detection model to the client satellite; The local training unit is used to train the target detection model locally using locally stored remote sensing data to obtain the corresponding update gradient. The dynamic gradient compression processing unit is used to calculate the inter-satellite link quality, select a strategy based on the inter-satellite link quality and a preset compression method, select a compression method to compress the updated gradient, obtain the compressed gradient, calculate the compression error, and send the compressed gradient and corresponding metadata to the main satellite according to the aggregation command. Local cache units are used to store compression errors; The client selects the scheduling unit, which is used to select client satellites to participate in the federated aggregation in this round based on the status information of the client satellites and a preset selection strategy, and sends aggregation instructions to the selected client satellites; The first model update processing unit is used to decompress the received compressed gradients, update the target detection model based on the decompressed gradients and a synchronous-asynchronous hybrid aggregation mechanism, and obtain the updated target detection model. The synchronous-asynchronous hybrid aggregation mechanism includes: synchronously aggregating gradients received within a specified time, asynchronously fusing gradients that arrive after a timeout but match based on time decay; and calculating the model difference between the target detection model before and after the update, and sending the model difference to all client satellites or selected client satellites participating in the next round of federated aggregation. The second model update processing unit is used to update the target detection model based on the received model difference.

[0070] Furthermore, in this embodiment of the invention, the local training unit is also used to obtain the compression error stored locally, update the momentum accumulation error according to the compression error, and use the momentum accumulation error to correct the update gradient obtained by local training.

[0071] The above-mentioned units are hardware structures corresponding to the above-mentioned method steps. The specific working principle and beneficial effects of each unit can be found in the above-mentioned method, and will not be repeated here.

[0072] It should be noted that, in this document, relational terms such as “first” and “second” are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A communication-efficient federated object detection method for low earth orbit satellite constellation, characterized in that, include: Load and initialize the target detection model on the master satellite, and distribute the target detection model to the client satellites; The target detection model is trained locally on the client satellite using locally stored remote sensing data to obtain the corresponding update gradient; The inter-satellite link quality is calculated on the client satellite. Based on the inter-satellite link quality and the preset compression method, a strategy is selected to compress the update gradient, the compressed gradient is obtained, and the compression error is calculated and stored. Based on the status information of the client satellites, the main satellite selects the client satellites to participate in the federated aggregation in this round according to the preset selection strategy, and sends the aggregation command to the selected client satellites so that the selected client satellites send the compression gradient and the corresponding metadata to the main satellite. The received compressed gradients are decompressed on the main satellite, and the target detection model is updated based on the decompressed gradients and the synchronous-asynchronous hybrid aggregation mechanism to obtain the updated target detection model. The synchronous-asynchronous hybrid aggregation mechanism includes: synchronously aggregating gradients received within a specified time and asynchronously fusing gradients that arrive after the timeout but match based on time decay. Calculate the model difference between the target detection model before and after the update, and send the model difference to all client satellites or selected client satellites participating in the next round of federated aggregation, so that the client satellites can update the target detection model according to the model difference and carry out the next round of target detection model training and update until the training requirements are met.

2. The communication-efficient federated object detection method for low earth satellite constellation of claim 1, wherein, The quality of the inter-satellite link is calculated in the following way: ； in, Indicates the quality of inter-satellite links. This represents the normalized value of the inter-satellite link bandwidth availability between the client satellite and the master satellite. This represents the normalized value of the round-trip time delay of the inter-satellite link between the client satellite and the master satellite. This indicates the packet loss rate of the inter-satellite link between the client satellite and the master satellite. The weight representing the availability of inter-satellite link bandwidth. The weights representing the round-trip delay of inter-satellite links. The weights represent the packet loss rate of inter-satellite links.

3. The efficient federated target detection method for low-Earth orbit satellite constellations according to claim 1, characterized in that, The preset compression method selection strategy includes: When the quality of the inter-satellite link is less than the first quality threshold, the gradient is compressed using the Top-K gradient sparsity method, and the K value in the Top-K gradient sparsity method is set to the first value. When the quality of the inter-satellite link is not less than the first quality threshold and not greater than the second quality threshold, the gradient is compressed using the Top-K gradient sparsification method, and the K value in the Top-K gradient sparsification method is set to the second value. When the quality of the inter-satellite link is greater than the second quality threshold, the gradient is compressed using the low-rank matrix factorization method, and the rank value in the low-rank matrix factorization method is set to the third value.

4. The efficient federated target detection method for low-Earth orbit satellite constellations according to claim 1, characterized in that, The compression error is calculated in the following way: ； in, Indicates compression error. This indicates that the gradient is being updated. This represents the compression gradient.

5. The efficient federated target detection method for low-Earth orbit satellite constellations according to claim 1, characterized in that, The preset selection strategy includes: From the client satellites whose status update interval does not exceed the preset time, select the client satellites whose remaining energy is not lower than the preset energy threshold, whose computing load is not higher than the preset load threshold, and whose available memory is not lower than the preset memory threshold; Based on the selected client satellites, and considering their remaining energy, computational load, and remote sensing data quality, a corresponding importance score is calculated. The selected client satellites are sorted from highest to lowest importance score, and the top P client satellites are selected as the client satellites participating in this round of federated aggregation, where P is a preset positive integer.

6. The efficient federated target detection method for low-Earth orbit satellite constellations according to claim 5, characterized in that, Calculate the corresponding importance score using the following formula: ； in, Indicates the importance score, This indicates the remaining energy of the client satellite. This indicates the computing load of the client satellite. This indicates the quality of the client's satellite remote sensing data. This represents the weight corresponding to the remaining energy. This indicates the weight corresponding to the calculated load. This represents the weight corresponding to the quality of remote sensing data.

7. The efficient federated target detection method for low-Earth orbit satellite constellations according to claim 1, characterized in that, The update of the target detection model based on the decompressed gradient and synchronous-asynchronous hybrid aggregation mechanism includes: Synchronously aggregate the gradients received within a specified time to obtain the synchronously aggregated gradients; The target detection model is updated based on the synchronously aggregated gradient; Calculate the corresponding decaying learning rate based on the delay time of the matched gradient that arrives after the timeout. The object detection model is updated again based on the gradients that arrive after the timeout and match, and their corresponding decaying learning rates.

8. The efficient federated target detection method for low-Earth orbit satellite constellations according to claim 7, characterized in that, The following formula is used to synchronously aggregate the gradients received within a specified time to obtain the synchronously aggregated gradients: ； Based on the synchronously aggregated gradient, the target detection model is updated using the following formula; ； Based on the delay time corresponding to the matched gradient that arrives after a timeout, the corresponding decaying learning rate is calculated using the following formula: ； Based on the timed-out and matched gradients and their corresponding decaying learning rates, the object detection model is updated again using the following formula: ； in, This indicates synchronous gradient aggregation. Indicates the first The decompressed gradient express The corresponding amount of training data, Indicates the number of client satellites used for synchronization aggregation. This represents the updated object detection model parameters. This represents the object detection model parameters before the update. Indicates the learning rate. This represents the decaying learning rate. Represents an exponential function. Indicates the attenuation coefficient. Indicates the delay time. This represents the updated object detection model parameters. This represents the gradient that has reached its timeout and is matched after decompression.

9. The efficient federated target detection method for low-Earth orbit satellite constellations according to claim 1, characterized in that, The method further includes: when training and updating the target detection model in the next round, after obtaining the corresponding update gradient, acquiring the locally stored compression error, updating the momentum accumulation error according to the compression error, and using the momentum accumulation error to correct the update gradient obtained by local training. The momentum accumulation error is updated using the following formula based on the compression error: ； in, This represents the cumulative momentum error. This represents the cumulative momentum error from the previous round. Indicates the momentum coefficient. Indicates compression error; The update gradient obtained from local training is corrected using the following formula: ； in, This indicates that the gradient is being updated. This represents the corrected updated gradient.

10. A highly efficient federated target detection system for low-Earth orbit satellite constellations, characterized in that, include: The model initialization unit, client selection and scheduling unit, and first model update processing unit are deployed on the main satellite; the local training unit, dynamic gradient compression processing unit, local caching unit, and second model update processing unit are deployed on the client satellite. The model initialization unit is used to load and initialize the target detection model, and distribute the target detection model to the client satellite; The local training unit is used to train the target detection model locally using locally stored remote sensing data to obtain the corresponding update gradient. The dynamic gradient compression processing unit is used to calculate the inter-satellite link quality, select a strategy based on the inter-satellite link quality and a preset compression method, select a compression method to compress the updated gradient to obtain the compressed gradient, calculate the compression error, and send the compressed gradient and corresponding metadata to the main satellite according to the aggregation instruction. The local cache unit is used to store compression errors; The client selection and scheduling unit is used to select client satellites participating in the federated aggregation in this round based on the status information of the client satellites and a preset selection strategy, and send aggregation instructions to the selected client satellites; The first model update processing unit is used to decompress the received compressed gradient, update the target detection model based on the decompressed gradient and the synchronous-asynchronous hybrid aggregation mechanism, and obtain the updated target detection model. The synchronous-asynchronous hybrid aggregation mechanism includes: synchronously aggregating the gradients received within a specified time, and asynchronously fusing the gradients that arrive after the timeout and match based on time decay. And the model difference used to calculate the target detection model before and after the update, and send the model difference to all client satellites or selected client satellites participating in the next round of federated aggregation; The second model update processing unit is used to update the target detection model based on the received model difference.

Images