A method, system, device, and storage medium for collaborative processing of satellite data.

Abstract

This application discloses a collaborative processing method, system, device, and storage medium for satellite data, belonging to the technical field of satellite edge computing technology. The collaborative processing method for satellite data includes: splitting a data processing model into K sub-models; deploying the K sub-models to a satellite cluster, such that the i-th sub-model is deployed on the i-th computing satellite of the satellite cluster; after a target satellite collects satellite data, sending the satellite data to the satellite cluster, so that all computing satellites in the satellite cluster collaboratively process the satellite data according to preset rules to obtain a data processing result; and using the K-th computing satellite to send the data processing result to a ground station. This application can reduce data transmission latency during data processing and improve the processing efficiency of satellite data.

Description

Technical Field

[0001] This application relates to the field of satellite edge computing technology, and in particular to a method, system, device and storage medium for collaborative processing of satellite data. Background Technology

[0002] Satellites (such as remote sensing satellites) generate a large amount of data during their operation. In traditional solutions, satellite data is transmitted to ground stations or the cloud for processing via satellite-to-ground links. However, satellite-to-ground links are affected by bandwidth, visibility window, and scheduling, resulting in long transmission times for satellite data, which is difficult to meet the time-sensitive requirements of scenarios such as disaster emergency response and continuous monitoring.

[0003] Therefore, how to reduce data transmission latency and improve the processing efficiency of satellite data is a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0004] The purpose of this application is to provide a collaborative processing method, system, device, and storage medium for satellite data, which can reduce data transmission latency during data processing and improve the processing efficiency of satellite data.

[0005] To address the aforementioned technical problems, this application provides a collaborative processing method for satellite data, the method comprising:

[0006] The data processing model is split into K sub-models, and the K sub-models are deployed to a satellite cluster, such that the i-th sub-model is deployed on the i-th computing satellite of the satellite cluster; wherein the satellite cluster is a cluster including K computing satellites, 1≤i≤K;

[0007] After the target satellite collects satellite data, it sends the satellite data to the satellite cluster, so that all computing satellites in the satellite cluster can collaboratively process the satellite data according to preset rules to obtain data processing results. The preset rules are as follows: the first computing satellite processes the satellite data using the first sub-model to obtain the compression result of the first intermediate feature; the j-th computing satellite processes the compression result of the (j-1)-th intermediate feature using the j-th sub-model to obtain the compression result of the j-th intermediate feature; and the k-th computing satellite processes the compression result of the (K-1)-th intermediate feature using the k-th sub-model to obtain the data processing result, where 2≤j≤K-1.

[0008] The data processing results are transmitted to the ground station using the Kth computing satellite.

[0009] Optionally, if the satellite cluster receives multiple frames of satellite data, all computing satellites collaboratively process the satellite data according to preset rules, including:

[0010] All computing satellites perform pipelined collaborative processing of the satellite data according to preset rules, so that adjacent frames of satellite data are executed out of time; wherein, when the j-th computing satellite processes the n-th frame of satellite data, the (j-1)-th computing satellite processes the (n+1)-th frame of satellite data in parallel, and the (j+1)-th computing satellite processes the (n-1)-th frame of satellite data in parallel.

[0011] Optionally, after deploying the K-segment model to the satellite constellation, the method further includes:

[0012] A compression coding module is set up in the first computing satellite;

[0013] The compression encoding module and the decoding reconstruction module are configured in the j-th computing satellite;

[0014] The decoding and reconstruction module is installed in the Kth computational satellite;

[0015] The compression encoding module is used to compress and encode the intermediate features output by the local sub-model; the decoding and reconstruction module is used to decode the compression result of the intermediate features transmitted by the previous satellite to obtain reconstructed features, so that the local sub-model can process the reconstructed features.

[0016] Optionally, the intermediate features output by the local sub-model are compressed and encoded, including:

[0017] Sparsity selection is performed on the intermediate features output by the local sub-model to obtain sparse features;

[0018] The sparse features are subjected to bit quantization of a preset width to obtain a non-zero codeword sequence and quantization parameters;

[0019] Entropy encoding is performed on the non-zero codeword sequence based on the quantization parameters to obtain the compressed result of the intermediate features output by the local sub-model.

[0020] Optionally, sparsification can be performed on intermediate features output by the local sub-model, including:

[0021] The intermediate features output by the local sub-model are sparsified by using the mask generation module to obtain sparse features;

[0022] The training process of the mask generation module includes:

[0023] An initial mask for the input features is generated using the mask generator of the mask generation module, and noise is added to the initial mask.

[0024] Binarize the initial mask after adding noise to obtain a hard mask;

[0025] The input features are filtered using the hard mask to obtain the target sparse features;

[0026] Calculate the loss function value based on the target sparse features;

[0027] Backpropagation is performed using the loss function value, and the gradient of the binarization operation is set using the pass-through estimator during the backpropagation process to update the parameters of the mask generator.

[0028] Optionally, the data processing model can be split into K sub-models, including:

[0029] The data processing model is divided into K sub-models according to a preset granularity; wherein the preset granularity is at least one network layer or at least one network module.

[0030] This application also provides a collaborative processing method for satellite data, applied to the p-th computing satellite in a satellite constellation, wherein the satellite constellation includes K computing satellites, and the i-th computing satellite deploys the i-th sub-model of the data processing model, 1≤i≤K, 1≤p≤K-1. The collaborative processing method for satellite data includes:

[0031] After the local p-th sub-model outputs the p-th intermediate feature, the p-th intermediate feature is compressed and encoded.

[0032] The compressed result of the p-th intermediate feature is sent to the (p+1)-th computing satellite for processing;

[0033] The compression encoding operation for the p-th intermediate feature includes:

[0034] Sparsification selection is performed on the p-th intermediate feature to obtain sparse features;

[0035] The sparse features are subjected to bit quantization of a preset width to obtain a non-zero codeword sequence and quantization parameters;

[0036] Entropy encoding is performed on the non-zero codeword sequence based on the quantization parameters to obtain the compression result of the p-th intermediate feature.

[0037] This application also provides a collaborative processing system for satellite data, the system comprising:

[0038] The model deployment module is used to split the data processing model into K sub-models and deploy the K sub-models to the satellite cluster, so that the i-th sub-model is deployed on the i-th computing satellite of the satellite cluster; wherein the satellite cluster is a cluster including K computing satellites, 1≤i≤K;

[0039] The collaborative processing module is used to send the satellite data to the satellite cluster after the target satellite collects the satellite data, so that all computing satellites in the satellite cluster can collaboratively process the satellite data according to preset rules to obtain the data processing result. The preset rules are as follows: the first computing satellite processes the satellite data using the first sub-model to obtain the compression result of the first intermediate feature; the j-th computing satellite processes the compression result of the (j-1)-th intermediate feature using the j-th sub-model to obtain the compression result of the j-th intermediate feature; the k-th computing satellite processes the compression result of the (k-1)-th intermediate feature using the k-th sub-model to obtain the data processing result, where 2≤j≤K-1.

[0040] The result feedback module is used to send the data processing results to the ground station using the Kth computing satellite.

[0041] This application also provides a storage medium on which a computer program is stored, wherein the computer program, when executed, implements the steps of the above-described collaborative processing method for satellite data.

[0042] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor invokes the computer program in the memory to implement the steps of the above-described collaborative processing method for satellite data.

[0043] This application provides a collaborative processing method for satellite data. This method divides the data processing model into K sub-models and deploys these K sub-models to K computing satellites, such that the i-th sub-model is deployed on the i-th computing satellite of the satellite cluster. After the target satellite acquires satellite data, the K computing satellites in the satellite cluster use their locally deployed sub-models to collaboratively process the satellite data, and finally send the data processing results to the ground station. During the collaborative processing of satellite data by the computing satellites, the output of the previous sub-model serves as the input of the next sub-model, and the data transmitted between computing satellites is the compressed result of the intermediate features output by each sub-model, reducing the amount of data transmitted between computing satellites. Therefore, this application can reduce data transmission latency during data processing and improve the processing efficiency of satellite data. This application also provides a collaborative processing method for satellite data applied to the p-th computing satellite in a satellite cluster, a collaborative processing system for satellite data, a storage medium, and an electronic device, all with the above-mentioned beneficial effects, which will not be elaborated further here. Attached Figure Description

[0044] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A flowchart illustrating a collaborative processing method for satellite data provided in an embodiment of this application;

[0046] Figure 2 This is a schematic diagram of the structure of an on-orbit large-scale model collaborative inference system provided in an embodiment of this application;

[0047] Figure 3 A flowchart illustrating an on-orbit large-scale model collaborative inference method provided in this application embodiment;

[0048] Figure 4 A pipelined parallel timing diagram provided in an embodiment of this application;

[0049] Figure 5 This is a block diagram illustrating the principle of feature compression and transmission provided in an embodiment of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0051] Please see below. Figure 1 , Figure 1 This is a flowchart illustrating a collaborative processing method for satellite data provided in an embodiment of this application.

[0052] Specific steps may include:

[0053] S101: The data processing model is split into K sub-models, and the K sub-models are deployed to the satellite cluster so that the i-th sub-model is deployed on the i-th computing satellite of the satellite cluster.

[0054] This embodiment can be applied to the control end of a system containing a target satellite and a satellite constellation, so as to utilize the computing satellites in the satellite constellation to collaboratively process satellite data (such as remote sensing data) collected by the target satellite. The aforementioned satellite constellation is a cluster of K computing satellites, each equipped with a high-performance computing unit and possessing onboard data processing capabilities.

[0055] Before this step, a data processing model (e.g., a deep neural network model) can be determined based on the target satellite or the type of satellite data. After determining the data processing model, the number of computing satellites K in the satellite cluster can be determined, and the data processing model can be split into K sub-models. This step also deploys each sub-model to the satellite cluster, so that the i-th computing satellite in the cluster is equipped with the i-th sub-model, where i represents the sequence number of the computing satellite, 1≤i≤K. Through the above operations, each satellite can be responsible for only a specific stage of inference tasks, effectively avoiding the bottleneck of single-satellite computing power and improving the real-time performance of the satellite data processing process.

[0056] As a feasible implementation method, the process of splitting the data processing model into K-segment models can be as follows: splitting the data processing model into K-segment models according to a preset granularity; wherein, the preset granularity is at least one network layer or at least one network module.

[0057] S102: After the target satellite collects satellite data, it sends the satellite data to the satellite cluster so that all computing satellites in the satellite cluster can collaboratively process the satellite data according to preset rules to obtain data processing results.

[0058] After the target satellite collects satellite data, it can be controlled to send the satellite data to the satellite cluster. Specifically, the target satellite can send the satellite data to the first computing satellite, and the output result of the previous computing satellite becomes the input information for the next satellite, until the Kth computing satellite outputs the final data processing result.

[0059] Specifically, the preset rules followed by all computing satellites when collaboratively processing the satellite data are as follows: the first computing satellite processes the satellite data using the first sub-model to obtain the compression result of the first intermediate feature; the j-th computing satellite processes the compression result of the (j-1)-th intermediate feature using the j-th sub-model to obtain the compression result of the j-th intermediate feature; and the k-th computing satellite processes the compression result of the (K-1)-th intermediate feature using the k-th sub-model to obtain the data processing result, where 2≤j≤K-1.

[0060] It should be noted that after the intermediate features are generated by the local sub-models of the first to the (K-1)th computing satellites, compression operations can be performed on these intermediate features so that the compressed results can be transmitted to the next computing satellite. After the second to the Kth computing satellites receive the compressed results of the intermediate features transmitted by the previous computing satellite, they can first perform decompression operations on the compressed results and then process the decompressed results using their local sub-models.

[0061] S103: Use the Kth computing satellite to send the data processing results to the ground station.

[0062] In this context, the data processing result obtained after the Kth segment sub-model in the Kth computational satellite completes data processing is equivalent to the result obtained after inputting satellite data into the data processing model. This data processing result can include land cover type, forest fire detection results, sea surface temperature field data, etc.

[0063] This embodiment divides the data processing model into K sub-models and deploys these K sub-models to K computing satellites, such that the i-th sub-model is deployed on the i-th computing satellite of the satellite cluster. After the target satellite collects satellite data, the K computing satellites in the satellite cluster use their locally deployed sub-models to collaboratively process the satellite data and finally send the data processing results to the ground station. During the collaborative processing of satellite data by the computing satellites, the output of the previous sub-model serves as the input of the next sub-model, and the data transmitted between computing satellites is the compressed result of the intermediate features output by each sub-model, reducing the amount of data transmitted between computing satellites. Therefore, this embodiment can reduce data transmission latency during data processing and improve the processing efficiency of satellite data.

[0064] As for Figure 1 In a further description of the corresponding embodiment, if the satellite cluster receives multiple frames of satellite data, all computing satellites collaboratively process the satellite data according to preset rules, including:

[0065] All computing satellites perform pipelined collaborative processing of the satellite data according to preset rules, so that adjacent frames of satellite data are executed out of time; wherein, when the j-th computing satellite processes the n-th frame of satellite data, the (j-1)-th computing satellite processes the (n+1)-th frame of satellite data in parallel, and the (j+1)-th computing satellite processes the (n-1)-th frame of satellite data in parallel.

[0066] In the above process, there is a fixed time misalignment when adjacent satellites process different frames of data: when the j-th computing satellite is processing the n-th frame of satellite data, the (j-1)-th computing satellite is processing the (n+1)-th frame of satellite data; while the (j+1)-th computing satellite is processing the (n-1)-th frame of satellite data. It should be noted that the term "n-th frame of satellite data" is a broad concept, including both the raw n-th frame data directly acquired by the target satellite and the intermediate features or compressed results corresponding to the n-th frame generated after processing by any preceding segment sub-model in the link. In other words, regardless of the stage of data processing, any raw data originating from the n-th frame is collectively referred to as "n-th frame of satellite data."

[0067] As for Figure 1In a further description of the corresponding embodiment, after the K-segment sub-model is deployed to the satellite cluster, relevant modules for compression and decompression can be set in each computing satellite. For example, a compression encoding module can be set in the first computing satellite; the compression encoding module and the decoding and reconstruction module can be set in the j-th computing satellite; and the decoding and reconstruction module can be set in the K-th computing satellite.

[0068] The compression encoding module is used to compress and encode the intermediate features output by the local sub-model; the decoding and reconstruction module is used to decode the compression result of the intermediate features transmitted by the previous satellite to obtain reconstructed features, so that the local sub-model can process the reconstructed features.

[0069] Furthermore, the compression encoding module performs compression encoding on the intermediate features output by the local sub-model, which includes: performing sparsification selection on the intermediate features output by the local sub-model to obtain sparse features; performing bit quantization operation on the sparse features with a preset width to obtain a non-zero codeword sequence and quantization parameters; and performing entropy encoding on the non-zero codeword sequence based on the quantization parameters to obtain the compression result of the intermediate features output by the local sub-model.

[0070] The process of sparsification selection of intermediate features output by the local sub-model by the compression encoding module mentioned above includes: using the mask generation module to sparsify the intermediate features output by the local sub-model to obtain sparse features.

[0071] The training process of the mask generation module includes: generating an initial mask of input features using the mask generator of the mask generation module; adding noise to the initial mask; binarizing the initial mask with added noise (i.e., the soft mask) to obtain a hard mask; filtering the input features using the hard mask to obtain target sparse features; calculating the loss function value based on the target sparse features; performing backpropagation using the loss function value, and setting the gradient of the binarization operation using the pass-through estimator during the backpropagation process to update the parameters of the mask generator.

[0072] During backpropagation, the gradient of the binarization function is zero; the pass-through estimator ignores this property and directly passes the upstream gradient to the input, so that the weights can be updated normally based on the values ​​before the sign function.

[0073] This application also provides a collaborative processing method for satellite data applied to the p-th computing satellite, wherein the p-th computing satellite is... Figure 1The corresponding embodiment describes a satellite cluster containing K computing satellites. The i-th computing satellite deploys the i-th sub-model of a data processing model, where 1 ≤ i ≤ K and 1 ≤ p ≤ K-1. The collaborative processing method for the satellite data of the p-th computing satellite includes: after outputting the p-th intermediate feature from the local p-th sub-model, performing compression encoding on the p-th intermediate feature; and sending the compressed result of the p-th intermediate feature to the (p+1)-th computing satellite for processing.

[0074] The process of the Pth computational satellite compressing and encoding the pth intermediate feature includes: performing sparsification selection on the pth intermediate feature to obtain a sparse feature; performing bit quantization operation of the sparse feature with a preset width to obtain a non-zero codeword sequence and quantization parameters; and performing entropy encoding on the non-zero codeword sequence based on the quantization parameters to obtain the compression result of the pth intermediate feature.

[0075] In the above scheme, the data transmitted between satellites is not intermediate features, but compressed results of intermediate features, which can reduce the communication volume of inter-satellite links without reducing inference accuracy.

[0076] The process described in the above embodiments is illustrated below through practical applications of remote sensing satellites.

[0077] Remote sensing satellites generate high-resolution, large-scale images / sequence data during Earth observation missions. The traditional engineering workflow typically involves: satellite imaging, downlinking of raw images via a satellite-to-ground link, deep learning interpretation at ground stations / clouds, and output of results. However, the satellite-to-ground link is affected by bandwidth, visibility windows, and scheduling, resulting in long downlink times for raw images, making it difficult to meet the time-sensitive requirements of scenarios such as disaster emergency response and continuous monitoring. Therefore, related technologies are gradually shifting towards on-orbit processing or space edge computing: completing screening, identification, classification, and detection on the satellite as much as possible, and only downlinking results or summary information to reduce the amount of data transmitted back and improve near real-time service capabilities. Currently, single-satellite autonomous processing and constellation collaborative processing are important development directions for on-orbit intelligent processing.

[0078] As the parameters of deep neural networks (DNNs) gradually increase, especially large models based on Transformer Blocks, the requirements for computing and GPU / storage resources are high. Given the constraints of power consumption, heat dissipation, and radiation-resistant components on existing satellites, it is often difficult to directly deploy complete large models on a single satellite. Therefore, the following two approaches have emerged:

[0079] (1) Lightweight inference for single stars: The model is reduced in size and redeployed through distillation, pruning, quantization and other methods;

[0080] (2) Collaborative / Distributed Inference: The network is divided into layers / modules and executed on different nodes. Intermediate features / intermediate activations are passed between nodes to complete end-to-end inference.

[0081] "Segmenting the network and transmitting intermediate features" is a classic approach in the field of mobile-edge / cloud collaborative inference: some studies have decided at the layer granularity to run the front and back parts of the DNN on mobile devices and data centers / edges respectively to optimize latency / energy consumption; the core of this approach is to transmit intermediate features at the segmentation points. However, this approach also brings the following problems: the amount of intermediate feature data within a layer may be larger than the original input, making communication a bottleneck. In addition, in multi-accelerator / multi-machine scenarios, some studies have segmented the model into multiple stages in sequence and adopted pipeline parallelism to improve throughput and achieve communication and computation overlap, which is the basis of the engineering concept of "segmented deployment + pipeline".

[0082] The shortcomings of existing technologies and their causal chain can be summarized into the following three points, and the purpose of this solution is clarified accordingly: First, the limited space-to-ground link makes it difficult to achieve real-time "on-orbit acquisition - ground processing": The amount of raw data generated by remote sensing missions, especially high-resolution / high-precision image data, is huge, while the bandwidth of the space-to-ground link is limited and the available window is affected by the coverage of the orbit and ground station, resulting in long downlink time and serious queuing. As a result, the "downlink to the ground first and then inference" processing link cannot meet the needs of scenarios with extremely high timeliness requirements (such as disaster emergency response, continuous observation and change monitoring of ground targets, etc.), ultimately leading to the inability to achieve near real-time or even real-time remote sensing data processing. Therefore, one of the purposes of this embodiment is to move the inference as far forward as possible to complete it on-orbit, and only downlink the inference results / summary information, fundamentally reducing the amount of data transmitted back from space to ground and improving mission timeliness. Second, the limited computing power and resources of a single satellite make it difficult to deploy large models directly in orbit, while lightweighting causes a loss of accuracy: Existing satellite platforms are constrained by power consumption, heat dissipation, size, and the performance of radiation-resistant devices, resulting in limited available computing power and video / storage resources per satellite, making it difficult to directly load and run large models (such as large models based on Transformer). Existing alternatives often use lightweight deployment methods such as knowledge distillation, pruning, and quantization to adapt to resources, but these methods often sacrifice the expressive power of the model, leading to a decrease in inference accuracy and making it difficult to meet the requirements of high-precision remote sensing interpretation. Therefore, the second objective of this embodiment is to make large models feasible for operation in orbit without significantly reducing accuracy, and to improve the equivalent computing power and available memory through multi-node collaborative computing. Third, while model splitting can alleviate the "single-satellite undeployability" problem, it introduces a communication bottleneck for intermediate activations between segments: To overcome the lack of single-satellite resources, the existing or intuitive approach is to split the large model and deploy it in segments on multiple satellites / nodes for collaborative inference. However, for structures like Transformer Blocks, the intermediate activations that need to be transmitted between segments are usually high-dimensional tensors, and the amount of data increases rapidly with the sequence length and feature dimension. Direct transmission would significantly occupy the inter-satellite link bandwidth and introduce additional latency, thereby offsetting the benefits of pipelined parallelism and even limiting the overall system performance to communication rather than computation. Therefore, the third objective of this embodiment is to provide a low-overhead compression and transmission mechanism for intermediate activations between segments based on model splitting and pipelined parallelism. This mechanism significantly reduces the communication burden between segments while ensuring that the next segment can continue inference, thereby achieving near real-time collaborative inference of high-precision large models on multiple satellites / nodes.

[0083] It is evident that remote sensing images / sequence data possess the characteristics of "high resolution, high throughput, and strong timeliness." Traditional solutions typically involve completely downloading the original images to the ground before interpretation and computation. However, due to limitations in visibility windows, link bandwidth, and queuing scheduling, end-to-end processing time is prolonged, making it difficult to meet near-real-time requirements such as disaster emergency response and continuous observation. Moving inference to on-orbit also faces practical constraints: existing satellite platforms are limited by power consumption, size, heat dissipation, and radiation-resistant device capabilities. The available computing power and video / storage of a single satellite are insufficient to support large models based on structures such as Transformer Blocks. Therefore, common engineering practices include lightweight deployments such as knowledge distillation, pruning, and quantization to adapt to single-satellite resources. However, these often result in a decrease in model expressive power and a loss of accuracy, making it difficult to support high-precision remote sensing tasks.

[0084] To address the problems existing in the aforementioned related technologies, this embodiment provides a large-model collaborative inference scheme for low-Earth orbit satellite networks. This scheme can collaboratively deploy and run the same large model on multiple satellites / nodes under conditions of limited satellite-ground link (SGL) and inter-satellite link (ISL), and limited computing power and storage resources of a single satellite. It achieves high-precision, low-latency, pipelined parallel on-orbit inference while avoiding large communication overhead.

[0085] This embodiment provides the following overall reasoning framework for the segmented deployment and pipelined parallel reasoning of a constellation-oriented large-scale model in orbit: the same large model is divided into K segments according to the network layer and deployed on K computing satellites respectively. The reasoning is completed in orbit in a pipelined parallel manner, and only the final reasoning result or summary information is output and transmitted down, thereby improving timeliness and reducing the amount of data transmitted back under the condition of limited satellite-to-ground links.

[0086] This embodiment provides a joint compression mechanism of "sparserization / selection + low bit quantization + entropy coding" for intermediate features between segments: configurable compression of intermediate features between adjacent segments, and symmetric decoding and reconstruction at the receiving end to continue the next segment inference, thereby achieving a significant reduction in inter-satellite link communication traffic while maintaining inference accuracy that meets the threshold.

[0087] The following section introduces the overall system structure and data flow of large-scale collaborative reasoning:

[0088] Please see Figure 2 , Figure 2This is a schematic diagram of an on-orbit large-model collaborative inference system provided in an embodiment of this application. After the sensing satellite (i.e., the target satellite mentioned above) acquires / preprocesses images, it transmits the images to the first sub-model in computing satellite 1 for processing to obtain intermediate features. The intermediate features are then compressed and transmitted to the second sub-model in computing satellite 2 to obtain intermediate features. The intermediate features are then compressed and transmitted to the next computing satellite, and so on. The Kth sub-model in computing satellite K outputs the inference result and transmits the inference result to the ground station, which can receive the inference result. In the above process, data is transmitted between satellites via inter-satellite links (ISL), and between satellites and the ground via satellite-ground links (SGL).

[0089] exist Figure 2 In the system shown, the input node is responsible for data acquisition and preprocessing; computing nodes 1 to K respectively deploy the model segments 1 to K; a feature compression encoding / decoding reconstruction module is set up next to each segment link to compress intermediate features into a bit stream for transmission and restore them into a feature tensor that can continue to be reasoned at the receiving end. Figure 2 The arrows in the diagram indicate the direction of the data flow, in the following order: input data, segment 1 inference, intermediate features (compression), segment 2 inference, ..., segment K inference, and output result.

[0090] This scheme consists of input nodes (which can be sensing satellites or front-end nodes) and several computing nodes (which can be multiple computing satellites or multiple edge devices). The input node acquires the data to be processed (e.g., remote sensing images / sequences) and sends the data into the first sub-model. Each computing node carries only one network layer of the large model and collaboratively completes one end-to-end inference in sequence. Intermediate features are transmitted between adjacent segments via inter-satellite links or inter-node links, rather than transmitting raw data; finally, only the inference results or summary information are output to reduce the amount of data transmitted back.

[0091] The large model is segmented as follows: The large model to be deployed is segmented into K sub-models by network layers or network modules. The segmentation granularity can be several consecutive layers or several consecutive modules (e.g., a TransformerBlock). The input of each sub-model is the intermediate features output from the previous segment, and the output is the intermediate features required for the next segment, with the last segment outputting the final result.

[0092] After model segmentation, the k-th segment sub-model can be deployed on the k-th computing node. Each node locally stores only the parameters and cache required for its own segment, reducing the memory pressure on a single node. After deployment, each segment is numbered using a unified segment number (stage_id, segment identifier) ​​to ensure consistent inter-segment connectivity.

[0093] Please see Figure 3 , Figure 3 A flowchart of an on-orbit large-scale model collaborative inference method provided in this application embodiment includes the following steps:

[0094] S301: Split the model and deploy it in segments so that the large model can be split into K segments and deployed on K satellites.

[0095] This step involves splitting the large model into K sub-models based on preset splitting points, generating a segment configuration StageCfg[k] (e.g., segment number, layer range, input / output tensor format). This step can then distribute / load the k-th sub-model to the k-th computation node, completing model weight loading and runtime environment initialization.

[0096] S302: Acquire and preprocess sensing data from remote sensing satellites.

[0097] The preprocessing process includes: normalizing the size of the input image / sequence, organizing it into blocks or batches, and generating a data frame number frame_id.

[0098] S303: The computing satellite uses a pipelined parallel inference method and performs intermediate feature compression and transmission.

[0099] The pipelined parallel inference process includes: input data enters the first sub-model, compressed intermediate features are passed between sub-models via links, and finally the inference result is output in the Kth sub-model. The intermediate features in this step are sent via inter-satellite links after being sparsified, quantized, and encoded.

[0100] S304: Output the inference results, i.e., the results are transmitted to the ground station or used in orbit.

[0101] Specifically, to improve processing throughput and timeliness, this scheme employs a pipelined parallel computing acceleration process as follows: When the k-th sub-model is processing the i-th frame of data, the (k-1)-th sub-model can process the i+1-th frame of data in parallel, thus forming a steady state where multiple segments work simultaneously. The pipeline includes a startup phase (pipeline not yet full), a steady-state phase (segments continuously run in parallel), and a termination phase (segments are emptied one by one after input stops).

[0102] Please see Figure 4 , Figure 4 This application provides a pipelined parallel timing diagram, illustrating the first sub-model in satellite 1, the second sub-model in satellite 2, the Kth sub-model in satellite K, the inference execution flow, the intermediate data transmission flow, and the inference result transmission flow. 1~N' represent the time window for data processing and the sequence number of the data frame.

[0103] Figure 4In the diagram, the horizontal axis represents time, and the vertical axis represents the sub-models from segment 1 to segment K (corresponding to different computation nodes). Each segment's processing within a frame includes: inference computation (Compute), feature compression encoding (Encode), link transmission / reception (Transmit), feature decoding and reconstruction (Decode), and proceeding to the next segment's computation. During the steady-state phase, different frames are interleaved across different segments, thereby improving overall throughput. When link transmission and segment computation can be parallelized, computation-communication overlap can be further enhanced, improving acceleration.

[0104] The following is an explanation of the compression and transmission of intermediate features between segments:

[0105] In the segmented inference / segmented training architecture of this scheme, the transmission object between adjacent satellites is the intermediate feature output by the previous satellite model (also known as intermediate activation, intermediate feature tensor, or "token" in the large model). This intermediate feature typically has a three-dimensional structure of "batch size × sequence length × feature dimension" (e.g., batch size N, sequence length S, feature dimension D), resulting in large data volume and high bandwidth consumption. To reduce inter-segment communication overhead, reduce transmission latency, and maintain the availability of the next network segment, this scheme proposes a combined compressed link of "adaptive sparsity + quantization + entropy-guided coding," and performs symmetric decoding and reconstruction at the receiving end to ensure that subsequent network segments can continue inference or backpropagation training. The compressed link is one of the key implementations of this scheme, applicable to both inference scenarios and segmented training scenarios (including backpropagation gradient backpropagation).

[0106] Please see Figure 5 , Figure 5 The present application provides a feature compression and transmission principle block diagram. The operations performed on the intermediate features include feature sparsification, low-bit quantization, entropy encoding, and encapsulation to obtain the compression result. The compression result is transmitted via the inter-satellite link (ISL). After receiving the compression result, the satellite sequentially performs unpacking, decoding, inverse quantization, and feature reconstruction to obtain the reconstructed features.

[0107] The input for satellite feature computation is the intermediate feature F (i.e., the token) output from the previous large model. The feature sparsification process includes: selectively retaining features (e.g., selecting the top K features based on token / channel importance, Top-K), outputting the sparse feature F' and mask information (mask / index). The low-bit quantization process includes: representing continuous-value features in low-bit form (e.g., 8-bit integer quantization), outputting the quantized features and quantization parameters. The entropy encoding process includes: encoding the quantized symbol sequence for further compression (e.g., entropy encoding), outputting a bitstream. The packetization process includes: encapsulating the bitstream and necessary metadata into data packets before transmission.

[0108] The receiving end performs unpacking, decoding, dequantization, and mask reconstruction in reverse order to obtain the reconstructed feature F_rec that can be used in the next segment.

[0109] To ensure inter-node communication, inter-segment data packets should ideally include at least the following fields: frame_id (frame number), stage_id (segment number), seq_id (fragment number, if fragmented), payload_len (payload length), mask (mask information), quant (quantization parameter), and a CRC (Cyclic Redundancy Check) field. These fields are essential for project implementation; the specific encoding format can be determined by the implementer.

[0110] The compressed training scheme includes the following steps A1-A3:

[0111] Step A1: Gumbel mask sparsification / selection.

[0112] To reduce the transmission overhead of intermediate features between segments without significantly reducing inference accuracy, this embodiment outputs the intermediate feature tensor in the k-th segment. A learnable mask generation module is introduced to adaptively select activation elements with higher information content. This represents the range of values, which is related to the batch size N, sequence length S, and feature dimension D. The mask generation module can be implemented using a lightweight linear layer or a multi-layer perceptron (MLP) to process the input intermediate feature tensor. Output a logit (logarithmic) tensor of the same shape. Furthermore, the differentiability is approximately maintained during the training phase by using Gumbel-Sigmoid (an algorithm for differentiable sampling of binary discrete variables).

[0113] During the training phase, a soft mask is introduced for each element in the intermediate feature tensor. Sampling is performed, specifically by introducing Gumbel noise. , The noise is obtained by inverse transform sampling. To adaptively select activation elements with higher information content, a binary mask is generated. During training, the Gumbel-Sigmoid approximation is used to preserve differentiability. . This represents random noise that follows a uniform distribution. Indicates a standard uniform distribution. Indicates a soft mask. This represents the Sigmoid function (an activation function). Elements representing intermediate feature tensors This represents the temperature parameter.

[0114] During forward propagation, to achieve true "select / discard", the soft mask is binarized to obtain the hard mask: ; (·) indicates an indicator function. This indicates a hard mask.

[0115] This step can also employ a straight-through estimator (STE) during backpropagation. approximate This ensures that the mask module is trainable. Represents the gradient of the hard mask. This represents the gradient of the soft mask.

[0116] By using hard masks to filter features element by element, sparse features are obtained. To avoid the discarded positions interfering with the gradient, gradient separation (stopgrad) is used.

[0117] ;in, This indicates element-wise multiplication, and stopgrad(·) means that the values ​​are retained in the forward pass but the gradient is not propagated in the backward pass.

[0118] The above process only applies to The non-zero elements preserved by the mask are subjected to subsequent "quantization + encoding", and the "mask / index information + quantization value" are encapsulated together and sent to the next segment; the receiving end performs symmetric decoding and reconstruction and uses it as the input for the next segment to continue inference.

[0119] Throughout the training process, the loss function needs to be calculated to update the model parameters of the mask, resulting in a total loss. Loss due to mission and sparse regularization terms Composition, that is The sparse regularization term can be expressed as . The weight coefficients of the sparse regularization term are represented by S, the sequence length is represented by D, the feature dimension is represented by i and j, and the index variables are represented by j. The sparsity weights are used to control the compression strength (the lower the retention ratio, the smaller the transmitted bits). By backpropagation, the mask generation module parameters and the parameters of each segment of the model are updated simultaneously, enabling the system to achieve the desired sparsity while meeting the task accuracy requirements.

[0120] To ensure a stable training process, temperature parameters Annealing is used to gradually decrease the mask size from large to small, so that the mask gradually approximates the discrete hard selection from a probabilistic soft selection. For example:

[0121] ;

[0122] Where t is the current training round, and T is the total number of rounds. and These are the initial temperature and the minimum temperature, respectively, with max indicating the maximum value.

[0123] When deploying inference, Gumbel noise is no longer introduced; instead, it is directly derived from the elements. Generate a deterministic mask and perform "selection" according to the same masking rules. Quantification coding transmission "Decoding and reconstruction" ensures the stability and reproducibility of the on-orbit inference process, avoiding output jitter caused by random sampling.

[0124] The above operations can be used to sparsify the intermediate features output by the local sub-model, resulting in sparse features.

[0125] Step A2: Quantification.

[0126] To further reduce the number of bits transmitted between segments, this scheme only performs b-bit quantization (where b is the quantization bit width) on the valid non-zero elements retained in the mask M, obtaining a non-zero codeword sequence and quantization parameters. Quantization employs a "batch dynamic range estimation + signed uniform quantization" method, and symmetric inverse quantization reconstruction is performed at the receiving end to ensure that subsequent model segments can continue inference. Based on the mask... Obtain sparse features Only the set of elements that satisfy M=1 Quantization is performed, and no values ​​are sent at other positions (represented by a mask / index as "discard"). Represents a set of sparse feature indices. Represents coordinates in three-dimensional space. Indicates the position of the binary mask. The value of .

[0127] Within each mini-batch, the absolute values ​​of the valid elements are statistically analyzed. and To avoid the influence of outliers, quantile truncation or moving average can be used. The smallest absolute value representing a valid feature. Represents the maximum absolute value of the effective features. express The absolute value of the eigenvalue at a given location.

[0128] Using signed quantization, where 1 bit is used for the symbol and the remaining (b-1) bits are used for amplitude encoding, the quantization step size is: For each valid element Generate integer codewords , , where sign( ) is a symbolic function. This is a rounding operation. 0.5t represents the rounding compensation term.

[0129] Since rounding is not differentiable, the quantization operator can be approximated by a gradient using a straight-through estimator (STE) during the training phase, allowing the quantization module to be trained end-to-end. The deterministic quantization described above can be used directly during the inference phase, without the need for an STE.

[0130] For the inverse quantization operation, the receiving end uses the side information. With coding Reconstructed approximation: Combine with mask Write the reconstructed values ​​back to the corresponding positions to obtain the reconstructed features. As input for the next segment of the model. This represents the reconstructed approximation. This represents the original input value.

[0131] Step A3: Entropy-guided coding.

[0132] To further reduce inter-segment transmission overhead, this embodiment performs entropy-guided coding on the quantized non-zero codeword sequence, preferably using Huffman coding or arithmetic coding. This step is used both for actual compression and for estimating the compressible bit length to support rate control.

[0133] Let the set of non-zero codewords after quantization be... Its set of symbol values ​​is The sending end counts each symbol. experience frequency And calculate the information entropy: ;in It can be obtained from the current mini-batch histogram, or by using a moving average to improve stability.

[0134] If a fixed b bits are used to represent each non-zero codeword, then the length of the uncoded bits is... After employing entropy coding, its average code length can be approximately given by the entropy. Taking Huffman coding as an example, the expected code length... The sending end performs the following steps on the non-zero codeword sequence (and its corresponding mask / index information):

[0135] Construct symbolic distribution And generate a code table (e.g., a Huffman tree / codeword table).

[0136] Entropy encoding is performed on the codeword sequence to generate a bitstream payload.

[0137] Encapsulate and send necessary side information along with the payload, including: code table / code length table (or reconstructable distribution parameters), bit length b, and quantization side information (such as...). , , ) and mask / index information (or its compressed representation).

[0138] The above operations enable entropy encoding of the non-zero codeword sequence based on the quantization parameters, resulting in compressed intermediate features output by the local sub-model.

[0139] During the decoding process, the receiving end first parses the side information to recover the code table, then performs entropy decoding on the payload to obtain the codeword sequence, and writes it back to the corresponding position using a mask / index to complete subsequent dequantization and feature reconstruction. (Unloaded) or If it contains only a single symbol, you can skip coding and send an empty tag directly or use fixed-length coding.

[0140] The implementation method of the computing node device is as follows:

[0141] Each computing node (e.g., a computing satellite) can be implemented as an onboard computing device, including at least a processor, memory, a communication interface, and optional compression acceleration logic, used to execute the model inference of this segment and complete inter-segment feature compression and transmission. The device may include a CPU (Central Processing Unit) for task control and communication protocol processing; an accelerator (which may be a GPU, Graphics Processing Unit, or NPU) for inference computation of this segment; memory for storing model parameters and feature caches of this segment; communication transceiver circuitry for inter-segment link transmission and reception; and optional compression acceleration logic (e.g., FPGA, Field Programmable Gate Array) for accelerating compression operations such as sparsity / quantization / encoding. The data path is: receiving input, inference computation, feature compression, and packet transmission to the next segment.

[0142] Traditional "raw image downlink - ground-based centralized inference" links are limited by the bandwidth and visibility window of the Satellite-Ground Link (SGL), inevitably leading to long latency and insufficient real-time performance. This embodiment breaks down a large model into multiple segments and deploys them on multiple computing satellites / nodes, enabling inference to be completed in orbit, requiring only the downlink of final results or summary information. This fundamentally reduces the amount of data transmitted back from "high-precision image level" to "result level," significantly shortening end-to-end processing latency and improving the timeliness and on-orbit autonomy of tasks such as emergency monitoring and continuous observation. Furthermore, compared to single-satellite solutions that rely on lightweight methods like distillation / pruning, leading to decreased accuracy, this embodiment provides equivalent greater computing power and storage space through multi-node collaboration, allowing large models to run without excessive scaling. Therefore, it maintains high-precision interpretation capabilities while meeting on-board resource constraints. Furthermore, although model splitting introduces inter-segment intermediate activation (token) transmission overhead, this embodiment introduces an adaptive compression link of "learnable sparsity / selection + quantization + entropy coding" between segments, and performs symmetrical decoding and reconstruction at the receiving end, which significantly reduces the number of communication bits between segments and weakens the communication bottleneck. With the pipeline parallelism and computation-communication overlap mechanism, each segment can process continuous data frames in a steady state, improve system throughput and further reduce average response latency, thereby achieving the comprehensive effect of "high-precision large model running with low latency on multiple stars / multi-nodes".

[0143] This application provides a collaborative processing system for satellite data, which may include:

[0144] The model deployment module is used to split the data processing model into K sub-models and deploy the K sub-models to the satellite cluster, so that the i-th sub-model is deployed on the i-th computing satellite of the satellite cluster; wherein the satellite cluster is a cluster including K computing satellites, 1≤i≤K;

[0145] The collaborative processing module is used to send the satellite data to the satellite cluster after the target satellite collects the satellite data, so that all computing satellites in the satellite cluster can collaboratively process the satellite data according to preset rules to obtain the data processing result. The preset rules are as follows: the first computing satellite processes the satellite data using the first sub-model to obtain the compression result of the first intermediate feature; the j-th computing satellite processes the compression result of the (j-1)-th intermediate feature using the j-th sub-model to obtain the compression result of the j-th intermediate feature; the k-th computing satellite processes the compression result of the (k-1)-th intermediate feature using the k-th sub-model to obtain the data processing result, where 2≤j≤K-1.

[0146] The result feedback module is used to send the data processing results to the ground station using the Kth computing satellite.

[0147] This embodiment divides the data processing model into K sub-models and deploys these K sub-models to K computing satellites, such that the i-th sub-model is deployed on the i-th computing satellite of the satellite cluster. After the target satellite collects satellite data, the K computing satellites in the satellite cluster use their locally deployed sub-models to collaboratively process the satellite data and finally send the data processing results to the ground station. During the collaborative processing of satellite data by the computing satellites, the output of the previous sub-model serves as the input of the next sub-model, and the data transmitted between computing satellites is the compressed result of the intermediate features output by each sub-model, reducing the amount of data transmitted between computing satellites. Therefore, this embodiment can reduce data transmission latency during data processing and improve the processing efficiency of satellite data.

[0148] Optionally, if the satellite cluster receives multiple frames of satellite data, the process of all computing satellites coordinating the processing of the satellite data according to preset rules includes: all computing satellites performing pipelined coordinating processing of the satellite data according to preset rules, so that adjacent frames of satellite data are executed out of time; wherein, when the j-th computing satellite processes the n-th frame of satellite data, the (j-1)-th computing satellite processes the (n+1)-th frame of satellite data in parallel, and the (j+1)-th computing satellite processes the (n-1)-th frame of satellite data in parallel.

[0149] Furthermore, it also includes:

[0150] The module setting module is used to set up a compression encoding module in the first computing satellite after the K-segment sub-model is deployed to the satellite cluster; it is also used to set up the compression encoding module and the decoding and reconstruction module in the j-th computing satellite; and it is also used to set up the decoding and reconstruction module in the K-th computing satellite.

[0151] The compression encoding module is used to compress and encode the intermediate features output by the local sub-model; the decoding and reconstruction module is used to decode the compression result of the intermediate features transmitted by the previous satellite to obtain reconstructed features, so that the local sub-model can process the reconstructed features.

[0152] Furthermore, the compression encoding module performs compression encoding on the intermediate features output by the local sub-model, which includes: performing sparsification selection on the intermediate features output by the local sub-model to obtain sparse features; performing bit quantization operation on the sparse features with a preset width to obtain a non-zero codeword sequence and quantization parameters; and performing entropy encoding on the non-zero codeword sequence based on the quantization parameters to obtain the compression result of the intermediate features output by the local sub-model.

[0153] Furthermore, the process of sparsification selection of intermediate features output by the local sub-model by the compression encoding module includes: using the mask generation module to sparsify the intermediate features output by the local sub-model to obtain sparse features.

[0154] The training process of the mask generation module includes:

[0155] An initial mask for the input features is generated using the mask generator of the mask generation module, and noise is added to the initial mask.

[0156] Binarize the initial mask after adding noise to obtain a hard mask;

[0157] The input features are filtered using the hard mask to obtain the target sparse features;

[0158] Calculate the loss function value based on the target sparse features;

[0159] Backpropagation is performed using the loss function value, and the gradient of the binarization operation is set using the pass-through estimator during the backpropagation process to update the parameters of the mask generator.

[0160] Furthermore, the process by which the model deployment module splits the data processing model into K-segment sub-models includes: splitting the data processing model into K-segment sub-models according to a preset granularity; wherein the preset granularity is at least one network layer or at least one network module.

[0161] This application embodiment also provides a collaborative processing system for satellite data, applied to the p-th computing satellite in a satellite cluster, wherein the satellite cluster includes K computing satellites, and the i-th computing satellite deploys the i-th sub-model of the data processing model, 1≤i≤K, 1≤p≤K-1. The collaborative processing system for satellite data includes:

[0162] The encoding module is used to perform compression encoding on the p-th intermediate feature after the p-th segment sub-model outputs the p-th intermediate feature locally;

[0163] The data transmission module is used to send the compressed result of the p-th intermediate feature to the (p+1)-th computing satellite for processing;

[0164] The process of the encoding module performing compression encoding on the p-th intermediate feature includes: performing sparsification selection on the p-th intermediate feature to obtain a sparse feature; performing bit quantization operation on the sparse feature with a preset width to obtain a non-zero codeword sequence and quantization parameters; and performing entropy encoding on the non-zero codeword sequence based on the quantization parameters to obtain the compression result of the p-th intermediate feature.

[0165] Since the embodiments of the system part correspond to the embodiments of the method part, please refer to the description of the embodiments of the method part for the embodiments of the system part, and they will not be repeated here.

[0166] This application also provides a storage medium on which a computer program is stored, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0167] This application also provides an electronic device that may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the electronic device may also include various network interfaces, power supplies, and other components.

[0168] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

[0169] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only 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. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for collaborative processing of satellite data, characterized in that, include: The data processing model is split into K sub-models, and the K sub-models are deployed to a satellite cluster, such that the i-th sub-model is deployed on the i-th computing satellite of the satellite cluster; wherein the satellite cluster is a cluster including K computing satellites, 1≤i≤K; After the target satellite collects satellite data, it sends the satellite data to the satellite cluster, so that all computing satellites in the satellite cluster can collaboratively process the satellite data according to preset rules to obtain data processing results. The preset rules are as follows: the first computing satellite processes the satellite data using the first sub-model to obtain the compression result of the first intermediate feature; the j-th computing satellite processes the compression result of the (j-1)-th intermediate feature using the j-th sub-model to obtain the compression result of the j-th intermediate feature; and the k-th computing satellite processes the compression result of the (K-1)-th intermediate feature using the k-th sub-model to obtain the data processing result, where 2≤j≤K-1. The data processing results are transmitted to the ground station using the Kth computing satellite.

2. The collaborative processing method for satellite data according to claim 1, characterized in that, If the satellite cluster receives multiple frames of satellite data, all computing satellites collaboratively process the satellite data according to preset rules, including: All computing satellites perform pipelined collaborative processing of the satellite data according to preset rules, so that adjacent frames of satellite data are executed out of time; wherein, when the j-th computing satellite processes the n-th frame of satellite data, the (j-1)-th computing satellite processes the (n+1)-th frame of satellite data in parallel, and the (j+1)-th computing satellite processes the (n-1)-th frame of satellite data in parallel.

3. The collaborative processing method for satellite data according to claim 1, characterized in that, After deploying the K-segment sub-model to the satellite constellation, the following is also included: A compression coding module is set up in the first computing satellite; The compression encoding module and the decoding reconstruction module are configured in the j-th computing satellite; The decoding and reconstruction module is installed in the Kth computational satellite; The compression encoding module is used to compress and encode the intermediate features output by the local sub-model; the decoding and reconstruction module is used to decode the compression result of the intermediate features transmitted by the previous satellite to obtain reconstructed features, so that the local sub-model can process the reconstructed features.

4. The collaborative processing method for satellite data according to claim 3, characterized in that, The intermediate features output by the local sub-model are compressed and encoded, including: Sparsity selection is performed on the intermediate features output by the local sub-model to obtain sparse features; The sparse features are subjected to bit quantization of a preset width to obtain a non-zero codeword sequence and quantization parameters; Entropy encoding is performed on the non-zero codeword sequence based on the quantization parameters to obtain the compressed result of the intermediate features output by the local sub-model.

5. The collaborative processing method for satellite data according to claim 4, characterized in that, Sparsity selection is performed on the intermediate features output by the local sub-model, including: The intermediate features output by the local sub-model are sparsified by using the mask generation module to obtain sparse features; The training process of the mask generation module includes: An initial mask for the input features is generated using the mask generator of the mask generation module, and noise is added to the initial mask. Binarize the initial mask after adding noise to obtain a hard mask; The input features are filtered using the hard mask to obtain the target sparse features; Calculate the loss function value based on the target sparse features; Backpropagation is performed using the loss function value, and the gradient of the binarization operation is set using the pass-through estimator during the backpropagation process to update the parameters of the mask generator.

6. The collaborative processing method for satellite data according to claim 1, characterized in that, The data processing model is broken down into K-segment sub-models, including: The data processing model is divided into K sub-models according to a preset granularity; wherein the preset granularity is at least one network layer or at least one network module.

7. A method for collaborative processing of satellite data, characterized in that, The collaborative processing method for satellite data is applied to the p-th computing satellite in a satellite constellation, wherein the satellite constellation comprises K computing satellites, and the i-th computing satellite deploys the i-th sub-model of the data processing model, 1≤i≤K, 1≤p≤K-1. After the local p-th sub-model outputs the p-th intermediate feature, the p-th intermediate feature is compressed and encoded. The compressed result of the p-th intermediate feature is sent to the (p+1)-th computing satellite for processing; The compression encoding operation for the p-th intermediate feature includes: Sparsification selection is performed on the p-th intermediate feature to obtain sparse features; The sparse features are subjected to bit quantization of a preset width to obtain a non-zero codeword sequence and quantization parameters; Entropy encoding is performed on the non-zero codeword sequence based on the quantization parameters to obtain the compression result of the p-th intermediate feature.

8. A collaborative processing system for satellite data, characterized in that, include: The model deployment module is used to split the data processing model into K sub-models and deploy the K sub-models to the satellite cluster, so that the i-th sub-model is deployed on the i-th computing satellite of the satellite cluster; wherein the satellite cluster is a cluster including K computing satellites, 1≤i≤K; The collaborative processing module is used to send the satellite data to the satellite cluster after the target satellite collects the satellite data, so that all computing satellites in the satellite cluster can collaboratively process the satellite data according to preset rules to obtain the data processing result. The preset rules are as follows: the first computing satellite processes the satellite data using the first sub-model to obtain the compression result of the first intermediate feature; the j-th computing satellite processes the compression result of the (j-1)-th intermediate feature using the j-th sub-model to obtain the compression result of the j-th intermediate feature; the k-th computing satellite processes the compression result of the (k-1)-th intermediate feature using the k-th sub-model to obtain the data processing result, where 2≤j≤K-1. The result feedback module is used to send the data processing results to the ground station using the Kth computing satellite.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor invokes the computer program in the memory to implement the steps of the collaborative processing method for satellite data as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores computer-executable instructions, which, when loaded and executed by a processor, implement the steps of the collaborative processing method for satellite data as described in any one of claims 1 to 7.

Images