A method for task scheduling and on-orbit closed-loop processing of space-based edge computing based on multi-layer track cooperation

Abstract

The present application relates to the technical field of aerospace electronic information and space-based intelligent computing, and proposes a method for space-based edge computing task scheduling and on-orbit closed-loop processing based on multi-layer orbit cooperation, which constructs a low-orbit perception layer, a medium-orbit backbone layer, a high-orbit central layer and a low-altitude node layer to form a four-layer space-based edge computing cooperative architecture. The low-orbit perception layer satellite collects multi-load data and performs on-orbit fusion processing and target identification screening through the on-board GPU, uploads key data and discards regular data; the medium-orbit backbone layer receives the screened key data for regional multi-source correlation and autonomous decision-making or forwarding to the high-orbit; the high-orbit central layer performs global situation fusion and cross-regional task re-scheduling and issues instructions; the low-orbit satellite receives the instructions, executes attitude adjustment and load control and feeds back the status, forming a closed-loop iteration mechanism of collection, calculation, decision-making, execution and feedback. The response time delay is shortened from days to minutes, the bandwidth occupation is reduced by more than 90%, and the on-board autonomous decision-making capability and resource utilization efficiency are improved.

Description

Technical Field

[0001] This invention relates to the technical field of aerospace electronic information and space-based intelligent computing, specifically to a method for space-based edge computing task orchestration and on-orbit closed-loop processing based on multi-layer orbital collaboration. Background Technology

[0002] See Figure 1 Currently, satellite remote sensing and situational awareness systems generally adopt a working mode of "on-board acquisition—ground transmission—ground processing—decision uploading." Low Earth Orbit (LEO) remote sensing satellites carry various types of payloads, including optical, infrared, multispectral, and electromagnetic payloads. After acquiring raw data for Earth observation or space target sensing, massive amounts of data are transmitted to ground stations via satellite-to-ground links. Ground data processing centers then perform decoding, correction, target identification, and information extraction to generate decision information. Finally, commands are uploaded to the satellite for execution via telemetry, tracking, and command links. However, this approach has the following shortcomings: 1. The response time is extremely long, which cannot meet the real-time requirements. Due to the communication window limitation between the satellite and the ground station (the visibility time of each orbit of a low-orbit satellite is only 10 to 15 minutes), the massive amount of raw data (the daily data output of a single high-resolution remote sensing satellite can reach TB level) must wait for the ground station to pass over before it can be transmitted. Moreover, due to the bandwidth limitation of the satellite-ground link, the data transmission itself takes a long time.

[0003] Therefore, in the traditional model, the end-to-end latency from satellite acquisition to users obtaining useful information from remote sensing data is typically several hours or even days. According to publicly available statistics, the average time for existing systems from on-board imaging to ground processing and generation of decision-making information exceeds 6 hours. For time-sensitive tasks such as natural disaster monitoring, maritime target tracking, and military situational awareness, this "time-level" response speed is completely unacceptable.

[0004] 2. The bandwidth bottleneck of the satellite-to-ground link leads to a large amount of invalid data being transmitted back. Because the satellite lacks sufficient computing power for real-time on-orbit data processing and target identification, it does not have the ability to "filter data and transmit it back on demand".

[0005] Therefore, in the traditional mode, the raw data collected by remote sensing satellites contains a large amount of invalid information (such as cloud-covered areas and background areas with no target changes). However, traditional solutions lack effective on-board filtering capabilities, and all raw data must be transmitted down using valuable satellite-to-ground link bandwidth. According to statistics, in typical scenarios such as ocean monitoring, the proportion of valid target data is usually less than 5% of the total data collected, and a large amount of bandwidth is wasted on invalid data.

[0006] 3. The ground processing center is overloaded and difficult to scale up because all the raw data from the satellites are aggregated to the ground for centralized processing, which lacks effective utilization of the on-board distributed computing capabilities.

[0007] Therefore, in the traditional model, with the large-scale deployment of low-Earth orbit satellite constellations (the number of satellites in orbit has exceeded 10,000 and continues to grow), massive amounts of data are flooding into ground processing centers, resulting in a severe shortage of ground computing resources and continuously extending processing queue times. The ground-based centralized processing architecture faces a serious scalability bottleneck.

[0008] 4. The satellite lacks autonomous decision-making capabilities and is highly dependent on ground intervention. As the satellite only serves as a "data collector" and "command executor," it does not possess the closed-loop autonomous capability of "perception-computation-decision."

[0009] Therefore, in the traditional mode, when existing satellites encounter unforeseen circumstances during mission execution (such as cloud cover in the target area or space debris threats), they must wait for ground-based rescheduling instructions and cannot autonomously adjust their mission plans in orbit. This results in low mission execution efficiency, and in extreme situations such as satellite-to-ground link interruption, the satellite will completely lose its responsiveness. Summary of the Invention

[0010] This invention proposes a space-based edge computing task orchestration and on-orbit closed-loop processing method based on multi-level orbital collaboration. It mainly addresses the shortcomings of the traditional satellite remote sensing and situational awareness system's "on-board acquisition - ground transmission - ground processing - decision uploading" working mode, which has problems such as long response time (days), serious waste of satellite-ground link bandwidth, lack of on-board autonomous decision-making capability, and excessive load on ground processing center. As a result, it cannot meet the time-sensitive tasks such as disaster monitoring, maritime target tracking, and military situational awareness that require minute-level closed-loop response.

[0011] A method for orchestrating and processing space-based edge computing tasks based on multi-level orbital coordination, designed for this purpose, includes the following steps: Step 1: Low-Earth orbit (LEO) sensing layer satellites collect raw sensing data and perform on-orbit fusion processing and target identification and screening; Step 2: The core satellites in the middle orbit receive and filter key data and make regional decisions or forward the data. Step 3: The high-orbit central satellites perform global situational awareness integration and cross-regional mission reorganization and issue commands; Step 4: After receiving the decision-making instructions, the low-orbit sensing layer satellites execute the corresponding actions and feed back the execution status, forming a closed-loop iterative mechanism.

[0012] In step one, the low-orbit sensing layer satellite synchronously collects raw sensing data through multiple types of payloads, and the onboard GPU edge computing unit performs spatiotemporal registration and preliminary fusion of multimodal data to form a multidimensional sensing data tensor under a unified spatiotemporal reference. The payloads include optical cameras, infrared detectors, multispectral scanners, and electromagnetic signal detection equipment. Spatiotemporal registration and preliminary fusion are based on satellite high-precision ephemeris and atomic clock synchronization to unify optical, infrared, multispectral, and electromagnetic data into the WGS-84 geodetic coordinate system and the same time reference, and heterogeneous image registration is performed through feature matching algorithms.

[0013] In step one, the onboard GPU loads a lightweight deep learning model to perform real-time in-orbit inference on the fused perception data. The inference results include target detection results, change detection results, and anomaly event markers. Based on the inference results, data filtering is performed. Key data containing targets, changes, or anomalies are compressed and uploaded to the mid-orbit backbone layer, while regular data without targets or changes are compressed and stored or deleted and discarded in orbit.

[0014] In step one, the on-orbit real-time inference of the lightweight deep learning model undergoes a three-level screening process, and the key data blocks that pass the three-level screening are uploaded after compression. The first-level pixel-level filtering uses a semantic segmentation model to identify and mark invalid areas such as clouds, shadows, and ocean backgrounds. The second-level target screening runs the target detection model within the effective observation area and outputs the target bounding box, category, and confidence score. The third-level event-based filtering is based on comparing time-series data to calculate the changing areas and detect preset abnormal events.

[0015] In step two, after receiving key data uploaded by the low-Earth orbit (LEO) sensing layer, the medium-Earth orbit (MEO) backbone satellites perform data decompression and integrity verification. They also perform spatiotemporal nearest neighbor correlation of the sensing results from multiple LEO satellites within the same area, and fuse multi-view observations to improve target positioning accuracy and reduce positioning error. For regional-level missions, the MEO backbone satellites directly generate decision commands and send them to LEO satellites for execution, with regional-level decision delay controlled within the corresponding unit time. For global missions, the correlated regional situational data is forwarded to the high-Earth orbit (HEO) central layer via inter-satellite laser links.

[0016] In step three, after the high-orbit central layer satellites aggregate the sensing results uploaded by the regional mid-orbit nodes, they unify all data into the J2000 geocentric inertial coordinate system and UTC time reference. Based on the global target feature library, they associate the observation results of the same target in different regions, run Kalman filtering or particle filtering to predict the target trajectory within a preset prediction time, and generate a global situation map. According to the changes in the global situation and the macro-task instructions input by the ground cloud center, the high-orbit central layer satellites dynamically adjust the priority of observation tasks in each region, reallocate the observation resources of the low-orbit sensing layer satellites, generate a task re-orchestration instruction set, and send it to the relevant low-orbit satellites through the mid-orbit backbone layer. The high-orbit central layer uploads the global situation data to the ground cloud computing center via the Q / V band satellite-to-ground link.

[0017] In step three, the intermediate orbit backbone satellites use the labeled data accumulated during in-orbit operation to perform lightweight federated learning or incremental learning, continuously optimize model parameters, and upload the incremental model parameters to the high orbit central layer; the high orbit central layer collects the incremental model parameters of each intermediate orbit node, performs federated aggregation to generate a global model, and then distributes it to each intermediate orbit node and low orbit sensing layer satellite.

[0018] In step four, after receiving the decision command, the low-orbit sensing layer satellite performs attitude adjustment through the onboard attitude control system to optimize the pointing accuracy, adjusts the imaging mode parameters of the payload, and performs continuous tracking observation of specific targets. The operational status of low-Earth orbit sensing layer satellites includes attitude positioning, payload activation, and imaging completion status, which are fed back to the high-Earth orbit central layer in real time via inter-satellite links, forming a closed-loop iterative mechanism.

[0019] The low-Earth orbit (LEO) sensing layer satellites, medium-Earth orbit (MEO) backbone layer satellites, and high-Earth orbit (HEO) central layer satellites constitute a four-layer space-based edge computing collaborative architecture: LEO sensing layer, MEO backbone layer, HEO central layer, and low-altitude node layer. The LEO sensing layer consists of multiple remote sensing / perception satellites deployed in low Earth orbit, each carrying multiple types of payloads, onboard GPU edge computing units, and lightweight inference models. The MEO backbone layer consists of multiple relay / computing satellites deployed in medium Earth orbit, each carrying an edge computing server. The HEO central layer consists of multiple high-Earth orbit satellites deployed in geostationary orbit, configured with high-performance edge computing servers and a global situational database. The low-altitude node layer consists of low-altitude / ground edge nodes such as UAVs, buoys, and ground mobile terminals, which are connected to the space-based edge computing network via inter-satellite links or space-to-space links.

[0020] A distributed task allocation mechanism based on capability assessment is adopted: when a low-Earth orbit sensing layer satellite receives an observation task, the on-board edge computing unit assesses its own capability value. The calculation is based on Capability = α × Battery + β × Payload + γ × Bandwidth - δ × TaskLoad; Where Battery is the battery level, Payload is the payload availability, Bandwidth is the remaining bandwidth, TaskLoad is the current task load, and α, β, γ, and δ are dynamic weighting coefficients. If Capability ≥ Threshold, the satellite will autonomously execute the mission to achieve local closed loop. If Capability < Threshold, the satellite sends a coordination request to the medium-Earth orbit backbone layer via inter-satellite links, and the medium-Earth orbit nodes coordinate with other low-Earth orbit satellites to complete the task. For complex tasks that require cross-regional coordination, the high-Earth orbit central layer performs global task rescheduling.

[0021] This invention constructs a four-layer space-based edge computing collaborative architecture consisting of a "low-Earth orbit sensing layer, a mid-Earth orbit backbone layer, a high-Earth orbit central layer, and a low-altitude node layer," optimizing the traditional "sensing-backhaul-decision" mode of satellites into a "sensing-computing-decision" on-board closed-loop mode. This architecture offers the following advantages: 1. Response latency is reduced from days to minutes. In traditional solutions, remote sensing satellites undergo multiple stages from imaging to ground processing, including waiting for ground station transit, data downlink, ground processing, decision uplink, waiting for satellite transit, and command execution, resulting in a total latency of 624 hours. This invention utilizes onboard GPU edge computing deployed in the low-Earth orbit (LEO) sensing layer to complete target identification and preliminary decision-making in orbit; regional autonomous decision-making is achieved through the mid-Earth orbit (MEO) backbone layer; and only global tasks require processing through the high-Earth orbit (HEO) central layer. The end-to-end response latency is reduced to 110 minutes, meeting the minute-level response requirements of time-sensitive tasks such as disaster monitoring and maritime target tracking.

[0022] 2. The bandwidth usage of the satellite-to-ground link is reduced by 90% to 99%. In traditional solutions, all raw data collected by the satellite (including a large amount of invalid data such as cloud cover and background data without targets) needs to be transmitted down, with the daily data volume of a single high-resolution remote sensing satellite reaching TB levels. This invention performs three-level filtering on-board (pixel level—target level—event level), transmitting only key data containing targets, changes, and anomalies, reducing the amount of transmitted data to 1% to 10% of the original data. In typical scenarios such as ocean monitoring, the proportion of effective data can be reduced to below 1%, resulting in particularly significant bandwidth savings.

[0023] 3. The onboard autonomous decision-making capability is improved by more than 15 times. In traditional solutions, satellites basically lack autonomous decision-making capabilities, and more than 95% of mission decisions rely on ground-based processes. This invention achieves onboard autonomous closed-loop processing for more than 80% of missions through a three-level autonomous decision-making mechanism: local closed-loop at the low-Earth orbit sensing layer (instant response after target detection), regional closed-loop at the medium-Earth orbit backbone layer (autonomous decision-making for regional collaborative missions), and global closed-loop at the high-Earth orbit central layer (cross-regional mission rescheduling).

[0024] 4. Ground processing center load reduced by over 80%. In traditional solutions, all raw data from satellites is aggregated at the ground processing center. With the surge in the number of satellites, the ground processing center faces a severe scalability bottleneck. This invention completes data screening and preliminary processing in orbit, transmitting only key information and global situational data back to the ground, reducing the data processing load of the ground processing center to less than 20% of that in traditional solutions.

[0025] 5. The socio-economic effects are as follows: 5.1 Reduce satellite operating costs: By reducing invalid data backhaul, significantly reduce satellite-to-ground link leasing costs and ground station construction and maintenance costs; 5.2 Improve satellite resource utilization: By filtering on-orbit data, the bandwidth of the satellite-to-ground link can be freed up for the transmission of more effective data; minute-level response capability can strongly support key tasks such as disaster emergency response, maritime rights protection, and military situation awareness; 5.3 Enhance social security capabilities: Minute-level response capabilities can effectively support key tasks such as disaster emergency response, maritime rights protection, and military situation awareness; 5.4. Promote the development of the space-based computing industry: create market demand for links in the industrial chain such as spaceborne intelligent chips and on-orbit intelligent processing systems. Attached Figure Description

[0026] Figure 1 The diagram illustrates the traditional satellite remote sensing data processing workflow. After a low-Earth orbit (LEO) satellite collects raw data, it transmits the data to a ground station via a satellite-to-ground link. The ground station then forwards the data to a ground data processing center for processing, generating decision commands, which are then uploaded back to the satellite for execution. This process involves numerous steps, is time-consuming, and consumes a large amount of bandwidth.

[0027] Figure 2 This is a diagram of a four-layer space-based edge computing collaborative architecture according to an embodiment of the present invention.

[0028] Figure 3 This is a flowchart of an embodiment of the present invention, showing the on-orbit closed-loop process of perception, computation, and decision-making.

[0029] Figure 4 This is a flowchart illustrating task allocation and collaboration among multiple tracks according to an embodiment of the present invention.

[0030] Figure 5 This is a flowchart illustrating a three-level data filtering process on a satellite according to an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. In order to make the above-mentioned objects, features and advantages of this application more apparent and understandable, many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0032] See Figure 2 The four-layer space-based edge computing collaborative architecture includes a low-altitude node layer, a low-Earth orbit sensing layer, a mid-Earth orbit backbone layer, and a high-Earth orbit central layer.

[0033] The low-altitude node layer consists of edge access nodes such as drones and buoys; The low-Earth orbit sensing layer consists of multi-payload remote sensing satellites equipped with onboard GPU edge computing units; The intermediate orbit backbone consists of relay / computing satellites, used for data relay and regional decision-making; The high-orbit central layer consists of GEO satellites, which are used to undertake global situational integration and mission reprogramming.

[0034] The low-altitude node layer, low-Earth orbit sensing layer, medium-Earth orbit backbone layer, and high-Earth orbit central layer are interconnected via laser / Ka-band inter-satellite links.

[0035] In addition, the 4 low-altitude node layer can be expanded with an optional extension layer. The optional extension layer consists of low-altitude / ground edge nodes such as UAVs, buoys, and ground mobile terminals, which are connected to the space-based edge computing network through inter-satellite links or space-to-space links. As a supplement to the perception layer, it provides near real-time, high-precision local environmental data; as an extension of the execution layer, it directly receives and executes decision commands from the space-based system.

[0036] See Figure 3 The low-Earth orbit (LEO) sensing layer completes multi-payload data acquisition → on-board GPU fusion processing → target identification and screening → local decision execution; key information is relayed through the mid-Earth orbit backbone → high-Earth orbit central fusion → global situational awareness fusion → mission re-arrangement, forming a complete closed loop.

[0037] See Figure 4 After receiving a mission, a low-Earth orbit satellite first assesses its own capabilities. If its capabilities are sufficient, it executes the mission locally. If its capabilities are insufficient, it requests coordination from the medium-Earth orbit backbone layer, which coordinates multi-satellite collaboration within the region. For complex cross-regional missions, the high-Earth orbit central layer performs global rescheduling.

[0038] See Figure 5 The raw data undergoes pixel-level filtering (cloud / shadow / background removal), target-level filtering (target detection and recognition), and event-level filtering (abnormal event detection) in sequence. Key data that passes the three-level filtering is progressively transmitted back (metadata → slices → full-resolution data), while invalid data is discarded on-board.

[0039] Specifically: The Low Earth Orbit (LEO) sensing layer consists of multiple remote sensing / perception satellites deployed in low Earth orbit (500-1500km). Each satellite carries multiple types of payloads, an onboard GPU edge computing unit, a lightweight inference model, a storage module, and a communication module. The multiple types of payloads include optical cameras, infrared detectors, multispectral scanners, and electromagnetic signal detection equipment. The onboard GPU edge computing unit is equipped with a high-performance GPU chip with TOPS-level computing power (using high-performance chips, computing power of no less than 10 TOPS (INT8), power consumption controlled within 50W, and radiation-hardened design). The lightweight inference model is a lightweight deep learning model deployed for target detection, change detection, and image segmentation. The target detection model is based on YOLOv8-Nano or a similar lightweight network, with a model size of <50MB. The change detection model is based on a lightweight version of the Siamese network. The semantic segmentation model is used for cloud, shadow, and background recognition. The storage module has a solid-state storage capacity of no less than 2TB, used to cache raw data and model parameters; the communication module supports inter-satellite laser communication (rate no less than 5Gbps) and Ka-band satellite-to-ground communication.

[0040] The on-orbit processing flow for low-Earth orbit sensing layer satellites includes: 1. Simultaneous acquisition of raw sensing data (optical images, infrared thermal images, multispectral data, electromagnetic signals, etc.) by multiple payloads; 2. The onboard GPU performs real-time on-orbit fusion processing of multimodal data and runs lightweight AI models for target detection, recognition, and screening; 3. Only the detected target information, abnormal events, and key feature data (not all raw data) are transmitted back to the high-orbit central layer via the middle track backbone layer; 4. Perform on-board compression storage or directly discard regular data without a target area, greatly reducing the backhaul of invalid data.

[0041] The Medium Earth Orbit (MEO) backbone consists of multiple relay / computing satellites deployed in medium Earth orbit (8000~20000km). Each satellite carries an edge computing server and performs the following functions: Data relay and protocol conversion: Receive preprocessed results uploaded from the low-orbit sensing layer, perform protocol conversion, and then forward them to the high-orbit central layer; Cross-domain task routing and load balancing: Dynamically schedule computational tasks based on the load status and task priority of each low-Earth orbit satellite; Incremental training of on-orbit models: Utilize labeled data accumulated during on-orbit operation to perform lightweight federated learning or incremental learning, and continuously optimize the parameters of lightweight deep learning models; Regional autonomous decision-making: For collaborative observation tasks within a region, medium-orbit satellites can directly allocate tasks and generate instructions without waiting for decisions from the high-orbit central control.

[0042] Specifically, each medium-Earth orbit backbone satellite is equipped with an edge computing server, an inter-satellite laser communication terminal, a storage module, and a federated learning aggregation module. The edge computing server is a CPU+GPU heterogeneous computing architecture with a computing power of no less than 50 TOPS. The inter-satellite laser communication terminal has a speed of no less than 10Gbps and supports simultaneous communication with multiple beams. The storage module has a solid-state storage capacity of no less than 10TB. The federated learning aggregation module is used to support a lightweight federated learning framework for secure aggregation.

[0043] The High Orbit (GEO) core consists of multiple high-orbit satellites deployed in geostationary orbit (35,786 km), equipped with high-performance edge computing servers and a global situational database, and undertakes the following functions: Global Situation Integration: Aggregates perception results uploaded from the central backbone layer of various regions and performs global spatiotemporal registration and situation integration; Cross-regional mission rescheduling: Dynamically adjust the priority of observation missions and resource allocation schemes in each region according to changes in the global situation; Edge large model parameter aggregation: Collect incremental model parameters from each mid-track node, perform federated aggregation to generate a global model, and then distribute it to each node; Collaborate with the ground cloud center: Synchronize global situational data to the ground cloud computing center and receive macro-level mission instructions from the ground.

[0044] Specifically, each high-orbit central satellite is equipped with a high-performance edge computing server, a global situation database, and a high-speed satellite-to-ground communication terminal. The computing power of the high-performance edge computing server is no less than 200 TOPS. The global situation database stores global target feature databases, historical trajectory databases, and environmental background databases. The high-speed satellite-to-ground communication terminal supports Q / V band satellite-to-ground communication with a speed of no less than 100Mbps.

[0045] The core innovation of this invention lies in constructing a complete onboard "sensing-computing-decision" closed-loop process. The closed-loop process specifically includes the following steps: Step 1: On-orbit acquisition and fusion of multi-payload data (Low Earth Orbit Sensing Layer). Multiple payload types (optical, infrared, multispectral, electromagnetic) from LEO sensing layer satellites simultaneously acquire raw sensing data. The onboard GPU edge computing unit performs spatiotemporal registration and preliminary fusion of the multimodal data, forming a multidimensional sensing data tensor under a unified spatiotemporal reference.

[0046] Specifically, during the operation of the low-Earth orbit sensing layer satellite, multiple payloads initiate data acquisition according to preset mission plans or ground commands. Example of acquisition parameters: Optical camera: panchromatic resolution 0.5m, multispectral resolution 2m, swath width 60km; Infrared detector: Mid-wave infrared (3~5μm), resolution 10m; SAR: 1m resolution, 20km swath width.

[0047] The raw data acquired by each payload is transmitted to the GPU edge computing unit via the onboard high-speed bus. The GPU performs spatiotemporal registration of the multimodal data: Based on high-precision satellite ephemeris (position accuracy <10m) and atomic clock synchronization (synchronization accuracy <1μs), the data of each payload are unified to the WGS-84 geodetic coordinate system and the same time reference; Feature matching algorithms such as SIFT or SuperPoint are used for heterogeneous image registration, and the registration accuracy is better than 1 pixel.

[0048] Step 2: Onboard intelligent target recognition and data filtering (low-Earth orbit perception layer). The onboard GPU loads lightweight deep learning models (such as YOLO series target detection networks, change detection networks, etc.) to perform real-time inference on the fused perception data, and performs data filtering based on the inference results.

[0049] The inference results include: target detection results (target type, location, size, confidence level); change detection results (change areas compared to historical images); and anomaly event markers (events that meet preset anomaly rules).

[0050] Critical data (including target, change, and anomaly data blocks) → are compressed and uploaded to the mid-orbit backbone layer; routine data (background data without targets or changes) → are compressed and stored on-board or discarded, and are not transmitted back.

[0051] Specifically, the onboard GPU loads a lightweight deep learning model to perform on-orbit inference on the fused perceptual data: 1) Pixel-level filtering: Run a semantic segmentation model (such as the lightweight DeepLabV3+) to identify and label invalid areas such as clouds (areas with coverage >80%), shadows, and ocean backgrounds, and generate a mask of valid observation areas.

[0052] 2) Target-level screening: Run the target detection model within the effective observation area.

[0053] For optical / infrared data, run the YOLOv8-Nano target detection network to output target bounding boxes, categories (ships / aircraft / vehicles / oil spills, etc.), and confidence scores (threshold 0.5). For SAR data, a CFAR (Constant False Alarm Rate) detector combined with a lightweight CNN classifier is used to detect maritime ship targets. The electromagnetic data is processed using a spectrum feature matching algorithm to identify abnormal signal sources.

[0054] 3) Event-level filtering: based on time-series data comparison.

[0055] The system compares the current detection results with historical detection results (last 7 days) stored on-board; calculates the changed areas (new targets, target displacement, target disappearance); and detects preset abnormal events (such as course deviation > 500m, illegal anchoring, oil spill area > 100m², etc.). Data filtering decisions are made: data blocks that pass the three-level filtering (including targets, changes, or anomalies) are uploaded after being compressed using JPEG2000 or HEVC (compression ratio 10:1~50:1); data blocks that fail the filtering are automatically deleted after 30 days of on-board storage and are not uploaded back.

[0056] Step 3: Medium-Earth Orbit Backbone Relay and Regional Decision-Making (Medium-Earth Orbit Backbone Layer). Medium-Earth Orbit backbone layer satellites receive key data uploaded from the low-Earth Orbit sensing layer, perform data decompression and integrity verification, and correlate regional multi-source data (correlate the sensing results of multiple low-Earth Orbit satellites in the same region). For regional tasks (such as regional search and rescue, local disaster assessment), medium-Earth Orbit satellites directly generate decision commands and issue them to low-Earth Orbit satellites for execution, achieving regional second-level closed loop. For global tasks, data is forwarded to the high-Earth Orbit central layer via inter-satellite links.

[0057] Specifically, the mid-Earth orbit backbone satellites receive key data uploaded by the low-Earth orbit sensing layer: 1) Data reception and verification: Receive compressed data packets, perform CRC check and integrity check, and initiate retransmission requests for lost data blocks; 2) Regional-level multi-source data association: The sensing results uploaded by multiple low-orbit satellites in the same region (e.g., within a 500km×500km range) are associated, and the multi-source observations of the same target are associated based on the spatiotemporal nearest neighbor algorithm; the multi-view observations are fused to improve the target positioning accuracy (positioning error <100m).

[0058] 3) Regional autonomous decision-making (for regional missions): For regional search and rescue missions: Directly generate search and rescue guidance instructions based on the target location and send them to relevant low-orbit satellites to adjust the observation attitude and payload parameters; For local disaster assessment: Generate disaster reports (including the affected area, the status of key facilities, etc.) and transmit them directly to the regional emergency center; Regional decision-making latency is controlled within 10 seconds.

[0059] Step 4: High-orbit central hub global situation fusion and mission rescheduling (high-orbit central layer). The high-orbit central layer satellites gather the perception results uploaded by the medium-orbit nodes in various regions, perform global spatiotemporal benchmark unification and situation fusion, cross-regional target correlation and trajectory prediction, and dynamically adjust the priority of observation missions and resource allocation schemes in various regions according to changes in the global situation. The mission rescheduling instructions are then sent to the relevant low-orbit satellites for execution through the medium-orbit backbone layer.

[0060] Specifically, the high-orbit central-layer satellites aggregate the sensing results uploaded by the regional mid-orbit nodes: 1) Global spatiotemporal benchmark unification: unify all data to the J2000 geocentric inertial coordinate system and UTC time benchmark to eliminate coordinate system deviations of nodes in different regions.

[0061] 2) Global target association and trajectory prediction: Based on the global target feature library, the observation results of the same target in different regions are associated; Kalman filtering or particle filtering is run to predict the target trajectory, with a prediction time of 30 minutes to 6 hours; a global situation map is generated (including the position, heading, speed and type of all identified targets).

[0062] 3) Cross-regional mission re-orchestration: Based on changes in the global situation (such as cross-regional target movement or the emergence of sudden threats), dynamically adjust the priority of observation missions in each region; reallocate observation resources of low-Earth orbit sensing layer satellites (adjust observation areas, revisit cycles, and payload modes); generate mission re-orchestration instruction sets, and distribute them to relevant low-Earth orbit satellites through the medium-Earth orbit backbone layer.

[0063] 4) In coordination with the ground cloud center, global situational data (compressed) is uploaded to the ground cloud computing center via Q / V band satellite-to-ground link; receive macro-level mission instructions from the ground (such as adding key observation areas or inserting emergency missions); ground synchronization cycle: once every 30 minutes (regular) or in real time (emergency).

[0064] The mission execution feedback and closed-loop iteration mechanism involves the low-orbit sensing layer satellite receiving and executing decision-making instructions (attitude adjustment, payload control, target tracking, etc.) and feeding back the execution status to the high-orbit central layer in real time, forming a complete closed-loop iteration mechanism.

[0065] After receiving decision-making instructions, the low-Earth orbit sensing layer satellites execute them as follows: Attitude adjustment command → Executed by the onboard attitude control system, with pointing accuracy better than 0.01°; Load control command → Adjust camera integration time, gain, imaging mode and other parameters; Target tracking command → Perform continuous tracking and observation of a specific target.

[0066] The execution status (attitude in position, payload activation, imaging completion) is fed back in real time to the mid-orbit backbone layer and the high-orbit central layer, forming a complete closed-loop iteration. Typical value of the closed-loop cycle (from sensing to execution feedback): Local closed loop (within the low Earth orbit): <1 second; Regional closed loop (via the middle orbit layer): <10 seconds; Global closed loop (via high orbit): <10 minutes.

[0067] Typical application scenarios: Scenario 1: Real-time monitoring of naval targets. A low-Earth orbit (LEO) sensing layer satellite, while in orbit, simultaneously collects data on a specific sea area using optical and SAR payloads. The onboard GPU runs a ship detection model, completing data acquisition and target detection within 30 seconds, identifying three suspicious ships. After three levels of filtering, only metadata (approximately 2KB) such as target location, size, and heading is uploaded. Upon receiving this data, the mid-Earth orbit (MEO) backbone layer completes regional correlation within 5 seconds and generates a continuous tracking command for distribution. From target appearance to the start of continuous tracking, the entire process takes less than 60 seconds, whereas traditional solutions require waiting for the satellite to pass overhead and transmit data before ground processing, resulting in a latency of over 6 hours.

[0068] Scenario 2: Natural Disaster Emergency Response. After an earthquake, the high-orbit central control layer receives emergency commands from the ground and completes cross-regional task rescheduling within 10 seconds, simultaneously imaging five low-orbit satellites covering the disaster area. The low-orbit satellites perform cloud screening and change detection in orbit, transmitting only the change detection results of the affected area (approximately 5% of the original data) back to the high-orbit central control layer via the medium-orbit backbone layer. The high-orbit central control layer completes global situation fusion and generates a disaster assessment report within 3 minutes, improving efficiency by two orders of magnitude compared to traditional methods (>12 hours).

[0069] The key technical details of the above method steps are as follows: 1. The task allocation strategy among multiple orbits adopts a distributed task allocation mechanism based on capability assessment. When a low-Earth orbit sensing layer satellite receives an observation task, the on-board edge computing unit first assesses its own capability value: Capability = α × Battery + β × Payload + γ × Bandwidth - δ × TaskLoad.

[0070] Where Battery is the battery level (0~1), Payload is the payload availability (0~1), Bandwidth is the remaining bandwidth (0~1), TaskLoad is the current task load (0~1), and α, β, γ, and δ are dynamic weighting coefficients.

[0071] If Capability ≥ Threshold, the satellite will autonomously execute the mission (local closed loop). If Capability < Threshold, the satellite sends a coordination request to the medium-Earth orbit backbone layer via inter-satellite links, and the medium-Earth orbit node coordinates with other low-Earth orbit satellites to complete the task. For complex tasks requiring cross-regional collaboration, the high-orbit central layer performs global task re-orchestration.

[0072] 2. Data filtering rules for on-board model inference: On-board models of low-Earth orbit sensing layer satellites undergo multi-level filtering: Level 1: Pixel-level filtering - Based on deep learning semantic segmentation, identify clouds, shadows, and invalid background areas, and mark valid observation areas; Level 2: Target-level screening – Run the target detection model to identify targets of interest such as ships, aircraft, vehicles, and changing areas, and output the target bounding boxes and confidence scores; Level 3: Event-level filtering – Based on time-series data comparison, detect abnormal events (such as channel deviation, illegal berthing, oil spill, etc.) and trigger alarms.

[0073] Only data blocks that have passed the three-level screening (usually 1% to 10% of the original data volume) are uploaded after compression.

[0074] 3. Optimization method for backhaul data volume: Adopt a gradual data transmission strategy. The first level of feedback includes metadata of the target detection results (target type, coordinates, confidence level, timestamp), with a data volume of approximately KB. Second-level return (on demand): Compressed image slices of the target area, with a data volume of approximately MB; Level 3 backhaul (on demand): High-resolution raw data of the target area, approximately GB in size.

[0075] By using a three-level progressive backhaul, the bandwidth usage of the satellite-to-ground link is minimized while meeting mission requirements.

[0076] 4. The fusion and interpretation process of multimodal sensing data: The fusion of multi-payload data from low-orbit sensing layer satellites adopts a three-level architecture of spatiotemporal alignment, feature-level fusion, and decision-level fusion. Spatiotemporal alignment: Based on high-precision ephemeris and clock synchronization, optical, infrared, multispectral, and electromagnetic data are unified to the same spatiotemporal reference; Feature-level fusion: Extract feature vectors from each modality and perform cross-modal feature weighting fusion through an attention mechanism; Decision-level fusion: Independent reasoning results from each modality are integrated into a single decision using DS evidence theory or a voting mechanism to improve the accuracy of target identification.

[0077] Figures 2-5 A method for orchestrating and processing space-based edge computing tasks based on multi-level orbital coordination, comprising the following steps: Step 1: Low-Earth orbit (LEO) sensing layer satellites collect raw sensing data and perform on-orbit fusion processing and target identification and screening; Step 2: The core satellites in the middle orbit receive and filter key data and make regional decisions or forward the data. Step 3: The high-orbit central satellites perform global situational awareness integration and cross-regional mission reorganization and issue commands; Step 4: After receiving the decision-making instructions, the low-orbit sensing layer satellites execute the corresponding actions and feed back the execution status, forming a closed-loop iterative mechanism.

[0078] By constructing a complete on-board closed-loop iterative mechanism, the traditional open-loop long link of "on-board data acquisition - ground transmission - ground processing - decision uploading" is transformed into an on-board closed-loop short link of "sensing - computing - decision making". This solution eliminates the limitations of the communication window between the satellite and the ground station and the links of massive data waiting to be transmitted in the traditional solution, and reduces the end-to-end response latency from days (6~24 hours) to minutes at the architectural level.

[0079] In step one, the low-orbit sensing layer satellite synchronously collects raw sensing data through multiple types of payloads, and the onboard GPU edge computing unit performs spatiotemporal registration and preliminary fusion of multimodal data to form a multidimensional sensing data tensor under a unified spatiotemporal reference. The payloads include optical cameras, infrared detectors, multispectral scanners, and electromagnetic signal detection equipment. Spatiotemporal registration and preliminary fusion are based on satellite high-precision ephemeris and atomic clock synchronization to unify optical, infrared, multispectral, and electromagnetic data into the WGS-84 geodetic coordinate system and the same time reference, and heterogeneous image registration is performed through feature matching algorithms.

[0080] Synchronous acquisition of multiple payload types (optical, infrared, multispectral, SAR, electromagnetic) ensures strict temporal synchronization of multimodal data, avoiding spatiotemporal misalignment during fusion due to inconsistent acquisition times of different payloads. Based on high-precision ephemeris (position accuracy <10m) and atomic clock synchronization (synchronization accuracy <1μs), the data of each payload are unified to the WGS-84 geodetic coordinate system and the same time reference, providing an accurate spatiotemporal reference for subsequent feature-level fusion. Feature matching algorithms such as SIFT or SuperPoint are used for heterogeneous image registration, and the registration accuracy is controlled to be better than 1 pixel, ensuring accurate spatial correspondence of different modal data and effectively supporting the accuracy of subsequent cross-modal feature fusion.

[0081] In step one, the onboard GPU loads a lightweight deep learning model to perform real-time in-orbit inference on the fused perception data. The inference results include target detection results, change detection results, and anomaly event markers. Based on the inference results, data filtering is performed. Key data containing targets, changes, or anomalies are compressed and uploaded to the mid-orbit backbone layer, while regular data without targets or changes are compressed and stored or deleted and discarded in orbit.

[0082] By running lightweight target detection models such as YOLOv8-Nano on a spaceborne GPU, target identification and data filtering are completed in orbit. Only key data containing targets, changes, or anomalies are uploaded, while routine data without targets or changes is stored or deleted in orbit, fundamentally changing the inefficient mode of "downloading all raw data" in traditional solutions. This filtering mechanism reduces the amount of data transmitted back to 1%~10% of the original data volume and reduces the bandwidth usage of the space-to-ground link by 90%~99%, solving the problem of wasted space-to-ground link bandwidth in traditional solutions due to a large amount of invalid data (cloud cover, no target background, etc., in scenarios such as marine monitoring, the proportion of effective data is usually less than 5% of the total collected data).

[0083] In step one, the on-orbit real-time inference of the lightweight deep learning model undergoes a three-level screening process, and the key data blocks that pass the three-level screening are uploaded after compression. The first-level pixel-level filtering uses a semantic segmentation model to identify and label invalid regions such as clouds, shadows, and ocean backgrounds. The pixel-level filtering model and task run a semantic segmentation model (such as the lightweight DeepLabV3+). Specifically, it identifies and labels invalid regions such as clouds (areas with coverage > 80%), shadows, and ocean backgrounds, and generates a mask for the valid observation area.

[0084] The second-level target screening involves running the target detection model within the effective observation area and outputting the target bounding box, category, and confidence score; it also involves running the target detection model within the effective observation area. Specific operations for different payloads are as follows: Optical / Infrared Data: Run the YOLOv8-Nano network to output the target bounding box, category (ship / aircraft / vehicle / oil spill, etc.) and confidence score (threshold 0.5).

[0085] SAR data: Detecting maritime ship targets by running a CFAR (constant false alarm rate) detector combined with a lightweight CNN classifier.

[0086] Electromagnetic data: Run a spectrum feature matching algorithm to identify abnormal signal sources.

[0087] The third-level event-based filtering is based on time-series data comparison to calculate the changing areas and detect preset abnormal events. Specifically, the current detection results are compared with the historical detection results (last 7 days) stored on the satellite.

[0088] Detection content: Calculate the changed area (new target, target displacement, target disappearance), and detect preset abnormal events (such as channel deviation > 500m, illegal anchoring, oil spill area > 100m², etc.).

[0089] The three-tiered, progressive screening system is clearly structured and has a well-defined division of labor. Pixel-level screening removes invalid areas such as clouds and shadows, reducing the computational burden on subsequent detection. Target-level screening accurately identifies targets of interest (ships, aircraft, vehicles, oil spills, etc.) within the valid area. Event-level screening captures dynamic changes and abnormal events based on time-series comparisons (channel deviation > 500m, illegal anchoring, oil spill area > 100m²). This three-tiered screening ensures that invalid data is eliminated step by step and valid information is accurately extracted, maximizing data compression while maintaining detection recall. Combined with JPEG2000 or HEVC compression (compression ratio 10:1~50:1), the proportion of valid data can be reduced to below 1% in typical scenarios such as marine monitoring, resulting in significant bandwidth savings.

[0090] In step two, after receiving key data uploaded by the low-Earth orbit (LEO) sensing layer, the medium-Earth orbit (MEO) backbone satellites perform data decompression and integrity verification. They also perform spatiotemporal nearest neighbor correlation of the sensing results from multiple LEO satellites within the same area, and fuse multi-view observations to improve target positioning accuracy and reduce positioning error. For regional-level missions, the MEO backbone satellites directly generate decision commands and send them to LEO satellites for execution, with regional-level decision delay controlled within the corresponding unit time. For global missions, the correlated regional situational data is forwarded to the high-Earth orbit (HEO) central layer via inter-satellite laser links.

[0091] The medium-Earth orbit (MEO) backbone layer uses a spatiotemporal nearest neighbor algorithm to connect multi-source observation data from multiple low-Earth orbit (LEO) satellites in the same region (e.g., 500km × 500km), fusing multi-view observations to improve target positioning accuracy to <100m. More importantly, the MEO backbone layer possesses regional-level autonomous decision-making capabilities—for regional search and rescue missions, local disaster assessments, and other regional-level tasks, MEO satellites can directly generate search and rescue guidance commands or disaster reports and distribute them to LEO satellites and regional emergency centers, without waiting for decisions from high-Earth orbit or ground stations, keeping regional-level decision-making latency within 10 seconds. This represents a breakthrough from "day-level waiting" to "second-level response," significantly improving the timeliness of regional emergency missions.

[0092] In step three, after the high-orbit central layer satellites aggregate the sensing results uploaded by the regional mid-orbit nodes, they unify all data into the J2000 geocentric inertial coordinate system and UTC time reference. Based on the global target feature library, they associate the observation results of the same target in different regions, run Kalman filtering or particle filtering to predict the target trajectory within a preset prediction time, and generate a global situation map. According to the changes in the global situation and the macro-task instructions input by the ground cloud center, the high-orbit central layer satellites dynamically adjust the priority of observation tasks in each region, reallocate the observation resources of the low-orbit sensing layer satellites, generate a task re-orchestration instruction set, and send it to the relevant low-orbit satellites through the mid-orbit backbone layer. The high-orbit central layer uploads the global situation data to the ground cloud computing center via the Q / V band satellite-to-ground link.

[0093] The high-orbit central layer aggregates sensing results uploaded by low-orbit sensing layer satellites from various regions. By unifying the J2000 geocentric inertial coordinate system and the UTC time reference, coordinate system biases between regional nodes are eliminated. Based on a global target feature database, observations of the same target in different regions are correlated. Kalman filtering or particle filtering is used to predict the target's trajectory for the next 30 minutes to 6 hours, generating a global situation map. The high-orbit central layer dynamically adjusts the priority of observation tasks in each region and reallocates observation resources from low-orbit sensing layer satellites based on changes in the global situation (such as target cross-regional movement or the emergence of sudden threats) and macro-level mission instructions input from the ground cloud center, achieving globally optimal resource allocation across regions. The global closed-loop cycle is controlled within 10 minutes, representing a significant improvement over traditional methods.

[0094] In step three, the intermediate orbit backbone satellites use the labeled data accumulated during in-orbit operation to perform lightweight federated learning or incremental learning, continuously optimize model parameters, and upload the incremental model parameters to the high orbit central layer; the high orbit central layer collects the incremental model parameters of each intermediate orbit node, performs federated aggregation to generate a global model, and then distributes it to each intermediate orbit node and low orbit sensing layer satellite.

[0095] The intermediate orbit (MEO) backbone layer utilizes labeled data accumulated during on-orbit operation to perform lightweight federated learning or incremental learning. Each MEO node only uploads incremental model parameters, not the original data, thus protecting data privacy in each region and reducing inter-satellite transmission bandwidth consumption. The high-orbit central layer collects incremental parameters from each MEO node, performs federated aggregation to generate a global model, and then distributes the aggregated global model to each MEO node and the low-orbit sensing layer satellites. This mechanism enables continuous on-orbit optimization of model parameters. The knowledge learned by each node on local data is aggregated into a global model through federated aggregation and then fed back to the entire network, allowing all satellite models to adapt to changes in on-orbit operating conditions and drift in the distribution of accumulated data, maintaining target recognition accuracy during long-term on-orbit operation.

[0096] In step four, after receiving the decision command, the low-orbit sensing layer satellite performs attitude adjustment through the onboard attitude control system to optimize the pointing accuracy, adjusts the imaging mode parameters of the payload, and performs continuous tracking observation of specific targets. The operational status of low-Earth orbit sensing layer satellites includes attitude positioning, payload activation, and imaging completion status, which are fed back to the high-Earth orbit central layer in real time via inter-satellite links, forming a closed-loop iterative mechanism.

[0097] After receiving decision-making commands, the low-Earth orbit (LEO) sensing layer satellites perform attitude adjustments (pointing accuracy better than 0.01°) and payload control (adjusting integration time, gain, and imaging mode) through their onboard attitude control system. This enables continuous tracking and observation of specific targets, achieving an upgrade from "passively executing pre-set ground commands" to "actively adjusting based on onboard decisions." The execution status is fed back to the high-Earth orbit (HEO) central layer in real time via inter-satellite links, forming a complete closed-loop iteration. The clear division of the three-level closed-loop cycle (local <1 second, regional <10 seconds, global <10 minutes) ensures that each level of mission can achieve a response speed that matches the timeliness requirements. Through the three-level autonomous decision-making mechanism, more than 80% of missions are processed autonomously onboard, improving onboard autonomous decision-making capabilities compared to traditional solutions (more than 95% dependent on ground).

[0098] The low-Earth orbit (LEO) sensing layer satellites, medium-Earth orbit (MEO) backbone layer satellites, and high-Earth orbit (HEO) central layer satellites constitute a four-layer space-based edge computing collaborative architecture: LEO sensing layer, MEO backbone layer, HEO central layer, and low-altitude node layer. The LEO sensing layer consists of multiple remote sensing / perception satellites deployed in low Earth orbit, each carrying multiple types of payloads, onboard GPU edge computing units, and lightweight inference models. The MEO backbone layer consists of multiple relay / computing satellites deployed in medium Earth orbit, each carrying an edge computing server. The HEO central layer consists of multiple high-Earth orbit satellites deployed in geostationary orbit, configured with high-performance edge computing servers and a global situational database. The low-altitude node layer consists of low-altitude / ground edge nodes such as UAVs, buoys, and ground mobile terminals, which are connected to the space-based edge computing network via inter-satellite links or space-to-space links.

[0099] The four-layer space-based edge computing collaborative architecture moves computing power from ground data centers to the space-based edge (low-Earth orbit GPU ≥ 10 TOPS, medium-Earth orbit ≥ 50 TOPS, high-Earth orbit ≥ 200 TOPS). By completing data screening and preliminary processing on-orbit, only key information and global situational data are transmitted back to the ground. The data processing volume of the ground processing center is reduced to less than 20% of that of traditional solutions, effectively solving the scalability bottleneck faced by the ground-based centralized processing architecture (against the backdrop of the number of satellites in orbit exceeding 10,000 and continuing to grow). The low-altitude node layer, through edge nodes such as UAVs and buoys, serves as a supplement to the perception layer and an extension of the execution layer, expanding the perception coverage and command execution endpoint of the space-based edge computing network.

[0100] A distributed task allocation mechanism based on capability assessment is adopted: when a low-Earth orbit sensing layer satellite receives an observation task, the on-board edge computing unit assesses its own capability value. The calculation is based on Capability = α × Battery + β × Payload + γ × Bandwidth - δ × TaskLoad; Where Battery is the battery level, Payload is the payload availability, Bandwidth is the remaining bandwidth, TaskLoad is the current task load, and α, β, γ, and δ are dynamic weighting coefficients. If Capability ≥ Threshold, the satellite will autonomously execute the mission to achieve local closed loop. If Capability < Threshold, the satellite sends a coordination request to the medium-Earth orbit backbone layer via inter-satellite links, and the medium-Earth orbit nodes coordinate with other low-Earth orbit satellites to complete the task. For complex tasks that require cross-regional coordination, the high-Earth orbit central layer performs global task rescheduling.

[0101] The distributed task allocation mechanism based on capability assessment fully considers the dynamic and heterogeneous nature of satellite resources—battery power, payload availability, remaining bandwidth, and current mission load jointly determine whether a satellite is suitable for undertaking a new mission. When capability is sufficient, the satellite autonomously executes the mission to achieve local closed-loop (<1 second). When capability is insufficient, it requests coordination from the medium-Earth orbit backbone layer through inter-satellite links. The medium-Earth orbit nodes coordinate with other low-Earth orbit satellites in the region to complete the mission collaboratively. For complex missions requiring cross-regional coordination, the high-Earth orbit central layer performs global re-orchestration. This three-level coordination mechanism avoids the problem of "mission delays or failures due to insufficient capability of a single satellite," achieves efficient utilization and load balancing of space-based computing resources, and improves the system's mission completion rate and resource utilization.

[0102] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for space-based edge computing task orchestration and on-orbit closed-loop processing based on multi-layer orbital coordination, characterized in that, Includes the following steps: Step 1: Low-Earth orbit (LEO) sensing layer satellites collect raw sensing data and perform on-orbit fusion processing and target identification and screening; Step 2: The core satellites in the middle orbit receive and filter key data and make regional decisions or forward the data. Step 3: The high-orbit central satellites perform global situational awareness integration and cross-regional mission reorganization and issue commands; Step 4: After receiving the decision-making instructions, the low-orbit sensing layer satellites execute the corresponding actions and feed back the execution status, forming a closed-loop iterative mechanism.

2. The method for space-based edge computing task orchestration and on-orbit closed-loop processing based on multi-layer orbital coordination according to claim 1, characterized in that: In step one, the low-orbit sensing layer satellite synchronously collects raw sensing data through multiple types of payloads, and the onboard GPU edge computing unit performs spatiotemporal registration and preliminary fusion of multimodal data to form a multidimensional sensing data tensor under a unified spatiotemporal reference. The payloads include optical cameras, infrared detectors, multispectral scanners, and electromagnetic signal detection equipment. Spatiotemporal registration and preliminary fusion are based on satellite high-precision ephemeris and atomic clock synchronization to unify optical, infrared, multispectral, and electromagnetic data into the WGS-84 geodetic coordinate system and the same time reference, and heterogeneous image registration is performed through feature matching algorithms.

3. The method for orchestrating and on-orbit closed-loop processing of space-based edge computing tasks based on multi-layer orbital coordination according to claim 1, characterized in that: In step one, the onboard GPU loads a lightweight deep learning model to perform on-orbit real-time inference on the fused perception data. The inference results include target detection results, change detection results, and anomaly event markers. Based on the reasoning results, data filtering is performed. Key data containing targets, changes, or anomalies are compressed and uploaded to the middle orbit backbone layer, while routine data without targets or changes are compressed and stored on orbit or deleted and discarded.

4. The method for space-based edge computing task orchestration and on-orbit closed-loop processing based on multi-layer orbital coordination according to claim 3, characterized in that: In step one, the on-orbit real-time inference of the lightweight deep learning model undergoes a three-level screening process, and the key data blocks that pass the three-level screening are uploaded after compression. The first-level pixel-level filtering uses a semantic segmentation model to identify and mark invalid areas such as clouds, shadows, and ocean backgrounds. The second-level target screening runs the target detection model within the effective observation area and outputs the target bounding box, category, and confidence score. The third-level event-based filtering is based on comparing time-series data to calculate the changing areas and detect preset abnormal events.

5. The method for orchestrating and on-orbit closed-loop processing of space-based edge computing tasks based on multi-layer orbital coordination according to claim 1, characterized in that: In step two, after receiving key data uploaded by the low-Earth orbit sensing layer, the medium-Earth orbit backbone satellites perform data decompression and integrity verification, and perform spatiotemporal nearest neighbor correlation on the sensing results of multiple low-Earth orbit satellites in the same area, and fuse multi-view observations to improve target positioning accuracy and positioning error; for regional-level tasks, the medium-Earth orbit backbone satellites directly generate decision instructions and send them to the low-Earth orbit satellites for execution, and the regional-level decision delay is controlled within the corresponding unit time. For global missions, the correlated regional situational data is forwarded to the high-orbit central layer via inter-satellite laser links.

6. The method for orchestrating and on-orbit closed-loop processing of space-based edge computing tasks based on multi-layer orbital coordination according to claim 1, characterized in that: In step three, after the high-orbit central layer satellites aggregate the sensing results uploaded by the regional mid-orbit nodes, they unify all data into the J2000 geocentric inertial coordinate system and UTC time reference. Based on the global target feature library, they associate the observation results of the same target in different regions, run Kalman filtering or particle filtering to predict the target trajectory within a preset prediction time, and generate a global situation map. According to the changes in the global situation and the macro-task instructions input by the ground cloud center, the high-orbit central layer satellites dynamically adjust the priority of observation tasks in each region, reallocate the observation resources of the low-orbit sensing layer satellites, generate a task re-orchestration instruction set, and send it to the relevant low-orbit satellites through the mid-orbit backbone layer. The high-orbit central layer uploads the global situation data to the ground cloud computing center via the Q / V band satellite-to-ground link.

7. The method for space-based edge computing task orchestration and on-orbit closed-loop processing based on multi-layer orbital coordination according to claim 1, characterized in that: In step three, the intermediate orbit backbone satellites use the labeled data accumulated during in-orbit operation to perform lightweight federated learning or incremental learning, continuously optimize model parameters, and upload the incremental model parameters to the high orbit central layer; the high orbit central layer collects the incremental model parameters of each intermediate orbit node, performs federated aggregation to generate a global model, and then distributes it to each intermediate orbit node and low orbit sensing layer satellite.

8. The method for space-based edge computing task orchestration and on-orbit closed-loop processing based on multi-layer orbital coordination according to claim 1, characterized in that: In step four, after receiving the decision command, the low-orbit sensing layer satellite performs attitude adjustment through the onboard attitude control system to optimize the pointing accuracy, adjusts the imaging mode parameters of the payload, and performs continuous tracking observation of specific targets. The operational status of low-Earth orbit sensing layer satellites includes attitude positioning, payload activation, and imaging completion status, which are fed back to the high-Earth orbit central layer in real time via inter-satellite links, forming a closed-loop iterative mechanism.

9. The method for space-based edge computing task orchestration and on-orbit closed-loop processing based on multi-layer orbital coordination according to claim 1, characterized in that: The low-orbit sensing layer satellites, medium-orbit backbone layer satellites, and high-orbit central layer satellites constitute a four-layer space-based edge computing collaborative architecture: low-orbit sensing layer—medium-orbit backbone layer—high-orbit central layer—low-altitude node layer. The low-Earth orbit (LEO) sensing layer consists of multiple remote sensing / perception satellites deployed in LEO, each carrying multiple types of payloads, onboard GPU edge computing units, and lightweight inference models; the medium-Earth orbit (MEO) backbone layer consists of multiple relay / computing satellites deployed in MEO, each carrying an edge computing server; the high-Earth orbit (HEO) central layer consists of multiple high-Earth orbit (HEO) satellites deployed in geostationary orbit, configured with high-performance edge computing servers and a global situational database; and the low-altitude node layer consists of low-altitude / ground edge nodes such as UAVs, buoys, and ground mobile terminals, which are connected to the space-based edge computing network via inter-satellite links or space-to-space links.

10. The method for orchestrating and on-orbit closed-loop processing of space-based edge computing tasks based on multi-layer orbital coordination as described in claim 1, characterized in that: A distributed task allocation mechanism based on capability assessment is adopted: when a low-Earth orbit sensing layer satellite receives an observation task, the on-board edge computing unit assesses its own capability value. The calculation is based on Capability = α × Battery + β × Payload + γ × Bandwidth - δ × TaskLoad; Where Battery is the battery level, Payload is the payload availability, Bandwidth is the remaining bandwidth, TaskLoad is the current task load, and α, β, γ, and δ are dynamic weighting coefficients. If Capability ≥ Threshold, the satellite will autonomously execute the mission to achieve local closed loop. If Capability < Threshold, the satellite sends a coordination request to the medium-Earth orbit backbone layer via inter-satellite links, and the medium-Earth orbit nodes coordinate with other low-Earth orbit satellites to complete the task. For complex tasks that require cross-regional coordination, the high-Earth orbit central layer performs global task rescheduling.

Images