A distributed reasoning method and device for low-orbit satellite constellations

Abstract

The application provides a distributed reasoning method and device for a low-orbit satellite constellation. The method comprises the following steps: a satellite receiving a processing task divides the processing task into multiple subtasks according to a preset split node; multiple decision schemes are constructed based on the multiple subtasks and multiple satellites; the processing time consumed for completing each subtask is calculated based on the data packet size of the subtask and the processing resource size of the satellite allocated with the subtask, the transmission time delay of the subtask in the transmission between the satellites is calculated based on the transmission speed between the satellites, and the total time required for each decision scheme is calculated based on the processing time and the transmission time; a relative value function is calculated based on the sampling rate of the satellite receiving the processing task, and an incentive function is calculated based on the relative value function; a decision value corresponding to each decision scheme is calculated based on the incentive function and the total time required for each decision scheme, and a final decision scheme is screened from the multiple decision schemes based on the decision value.

Description

Technical Field

[0001] The present invention relates to the field of low-orbit satellite communication technology, and in particular to a distributed reasoning method and device for low-orbit satellite constellations. Background Art

[0002] With the widespread deployment of communications and network technologies such as 5G, and the gradual entry of Wi-Fi 7 into the horizons of users and industries, the rapid advancement of these terrestrial internet technologies and the ongoing development of aerospace technology are accelerating the expansion of the internet into outer space. The emergence of satellite networks has successfully further expanded network communications services. Leveraging satellite communications' abundant spectrum resources, extensive coverage, and minimal interference from terrestrial environmental factors, networks can be deployed to every corner of the Earth, achieving "base station globalization" and continuous global communications coverage, eliminating the impact of geographical factors such as high mountains and low valleys, thereby effectively breaking down communication barriers in society.

[0003] The emergence and widespread adoption of new technologies such as Software Defined Networking (SDN) and Network Function Virtualization (NFV) have driven the development of satellite network architectures. These technologies separate the control and data planes of satellite networks, optimize the distribution of satellite network architectures, and achieve a hybrid distributed and centralized network architecture. However, due to factors such as transmission costs, latency, and loss in satellite communications, low-Earth orbit (LEO) satellites have become the primary component of satellite communications. LEO satellites offer lower latency, comparable to terrestrial fiber optic networks. They also offer lower transmission loss, which is essential for achieving high data transmission rates. LEO satellites also offer greater reliability and can be distributed across multiple orbital planes, ensuring that the failure of any one or more satellites will not significantly impact the system. Furthermore, LEO constellation systems are relatively inexpensive and typically have multiple backup satellites in orbit to readily replace damaged ones. Taking all these factors into consideration, the establishment of large-scale LEO constellation networks has become the prevailing trend. Low-orbit satellite communication systems can utilize communication technologies such as cellular communication, spot beams, multiple access, and frequency reuse to support real-time or near-real-time data transmission such as online games and video calls. After integration with the ground communication backbone network, it may give rise to new application scenarios.

[0004] However, the important development trend of applying low-orbit satellites to artificial intelligence reasoning applications in existing technologies faces many challenges. The existing technology mainly applies low-orbit satellites to artificial intelligence reasoning applications by sending tasks from the ground to a certain satellite, and then a single satellite completes the computing task. However, a single satellite often prioritizes computing power and has difficulty completing large-scale computing tasks. Summary of the Invention

[0005] In view of this, an embodiment of the present invention provides a distributed reasoning method for low-orbit satellite constellations to eliminate or improve one or more defects in the prior art.

[0006] One aspect of the present invention provides a distributed reasoning method for a low-orbit satellite constellation, the method comprising the following steps:

[0007] Any satellite in the low-orbit satellite constellation receives the processing task, and the satellite that receives the processing task divides the processing task into multiple subtasks according to preset splitting nodes;

[0008] Construct multiple decision-making schemes based on multiple subtasks and multiple satellites, and each decision-making scheme has a different combination of multiple subtasks allocated to multiple satellites;

[0009] The processing time required to complete each subtask is calculated based on the subtask packet size and the processing resource size of the satellite to which the subtask is assigned. The transmission delay of the subtask between satellites is calculated based on the transmission speed between satellites. The total time required for each decision solution is calculated based on the processing time and transmission time.

[0010] Calculate a relative value function based on a sampling rate of a satellite receiving a processing task, and calculate an incentive function based on the relative value function;

[0011] A decision value corresponding to each decision solution is calculated based on the incentive function and the total time required for each decision solution, and a final decision solution is obtained by screening multiple decision solutions based on the decision value.

[0012] Adopting the above scheme, this scheme divides the entire processing task into multiple subtasks, and corresponds to multiple decision-making schemes for the allocation of subtasks. In the process of selecting the decision-making scheme, it can comprehensively consider multiple indicators such as the effectiveness of satellite data collection, task transmission delay and processing delay. For task offloading and resource scheduling decisions, a quantum evolutionary algorithm based on incentive function and processing delay is designed, which can improve the processing efficiency of subtasks on the assigned satellites. Multiple satellites jointly solve the processing tasks, avoiding the problem of insufficient computing power of a single satellite, ensuring the processing effectiveness of processing tasks in low-orbit constellations and reducing complexity.

[0013] In some embodiments of the present invention, the processing resource size includes the number of CPU cores, and the step of calculating the processing time consumed to complete each subtask based on the data packet size of the subtask and the processing resource size of the satellite to which the subtask is assigned includes:

[0014] The number of CPU cores of the satellite to which the subtask is assigned is calculated based on the data packet size of the subtask and the number of CPU cores of the satellite to which the subtask is assigned;

[0015] calculating a sub-task processing rate based on the required CPU core number of the satellite assigned with the sub-task and the task processing rate of a single core in the satellite assigned with the sub-task;

[0016] calculating a processing time consumed by the satellite assigned with the sub-task to complete the sub-task based on the sub-task processing rate and the data packet size of the sub-task.

[0017] In some embodiments of the present application, in the step of calculating the required CPU core number of the satellite assigned with the sub-task based on the data packet size of the sub-task and the CPU core number of the satellite assigned with the sub-task, the required CPU core number of the satellite assigned with the sub-task is calculated according to the following formula:

[0018]

[0019] wherein, l n,m (i) represents the required CPU core number of the satellite m to which the sub-task i is assigned by the satellite n receiving the processing task, C n (i) represents the data packet size of the sub-task received from the satellite n, L m represents the total core number of the satellite m, p i represents the transmission parameter of the sub-task i, [] represents rounding, a n,m (i) is a judgment parameter, a n,m (i) = 1 if the sub-task i is assigned to the satellite m by the satellite n, otherwise a n,m (i) = 0.

[0020] In some embodiments of the present application, in the step of calculating the sub-task processing rate based on the required CPU core number of the satellite assigned with the sub-task and the task processing rate of a single core in the satellite assigned with the sub-task, the sub-task processing rate is calculated according to the following formula:

[0021] f n,m (i) = f m · l n,m (i);

[0022] wherein, f n,m (i) represents the task processing rate of the sub-task i assigned to the satellite m by the satellite n, f m represents the task processing rate of a single core in the satellite m assigned with the sub-task i, l n,m (i) represents the required CPU core number of the satellite m to which the sub-task i is assigned by the satellite n receiving the processing task.

[0023] In some embodiments of the present invention, in the step of calculating the processing time consumed by the satellite assigned the subtask to complete the subtask based on the subtask processing rate and the data packet size of the subtask, the processing time consumed by the satellite assigned the subtask to complete the subtask is calculated according to the following formula:

[0024]

[0025] in, represents the processing time consumed by subtask i assigned by satellite n to satellite m on satellite m, C n (i) represents the size of the subtask data packet received from satellite n, σ represents the CPU size required by the CPU in satellite m to process each bit of subtask data, and f n,m (i) represents the task processing rate of subtask i assigned by satellite n to satellite m on satellite m, a n,m (i) is the judgment parameter. If subtask i is assigned to satellite m by satellite n, then a n,m (i)=1, otherwise a n,m (i)=0.

[0026] In some embodiments of the present invention, the step of calculating the transmission delay of the subtask between satellites based on the transmission speed between satellites includes:

[0027] Calculate the transmission speed between satellites based on the transmission gain between satellites and the bandwidth of the inter-satellite link;

[0028] The transmission delay between satellites is calculated based on the transmission speed between satellites and the packet size of the subtask.

[0029] In some embodiments of the present invention, in the step of calculating the transmission speed between satellites based on the transmission gain between satellites and the inter-satellite link bandwidth, the transmission speed between satellites is calculated according to the following formula:

[0030]

[0031] in, W represents the transmission speed of data from satellite u to satellite m. u,m represents the intersatellite link bandwidth between satellite u and satellite m, P u represents the transmission power of satellite u, h u,m represents the transmission gain from satellite u to satellite m, N z represents the noise parameter.

[0032] In some embodiments of the present invention, the noise parameter is a preset constant.

[0033] In some embodiments of the present invention, the transmission gain between satellites is calculated according to the following formula:

[0034]

[0035] Among them, h u,m represents the transmission gain from satellite u to satellite m, represents the transmission gain of satellite u, represents the receiving gain of satellite m, represents the free space path loss parameter between satellite u and satellite m;

[0036] The free space path loss parameter between satellite u and satellite m is calculated according to the following formula:

[0037]

[0038] in, represents the free space path loss parameter between satellite u and satellite m, d u,m represents the transmission distance between satellite u and satellite m, c represents the speed of light, and p u,m Indicates the communication frequency between satellite u and satellite m.

[0039] In some embodiments of the present invention, in the step of calculating the transmission delay between satellites based on the transmission speed between satellites and the data packet size of the subtask, the transmission delay between satellites is calculated according to the following formula:

[0040]

[0041] in, represents the transmission delay between satellite u performing subtask k and satellite m performing subtask i, C n (i) represents the size of the data packet received from satellite n for subtask i, represents the transmission speed of data from satellite u to satellite m, a n,m (i) and a n,u (k) are all judgment parameters. If subtask i is assigned to satellite m by satellite n, then a n,m (i)=1, otherwise a n,m (i) = 0, if subtask k is assigned to satellite u by satellite n, then a n,u (k)=1, otherwise a n,u (k)=0.

[0042] In some embodiments of the present invention, the step of calculating the total time required for each decision solution based on the processing time and transmission time includes:

[0043] Calculate the end time parameter of each subtask in the decision solution based on the processing time and the transmission time;

[0044] The maximum end time parameter is used as the total time parameter required for the decision-making plan.

[0045] In some embodiments of the present invention, in the step of calculating the end time parameter of each subtask in the decision solution based on the processing time and the transmission time, the end time parameter of each subtask is calculated according to the following formula:

[0046]

[0047] in, represents the end time parameter of subtask i executed on satellite m, The end time parameter of subtask k executed on satellite u, represents the transmission delay between satellite u performing subtask k and satellite m performing subtask i, represents the processing time consumed by subtask i assigned by satellite n to satellite m on satellite m, and subtask k is the subtask executed before subtask i. If subtask i includes multiple previous subtasks, the maximum value in [] is taken.

[0048] In some embodiments of the present invention, in the steps of calculating the relative value function based on the sampling rate of the satellite receiving the processing task and calculating the incentive function based on the relative value function, the relative value function is calculated according to the following formula:

[0049] η n =α sp ;

[0050] Among them, η n represents the relative value function of the satellite n that receives the processing task, s sp represents the sampling rate of the satellite receiving the processing task, and α is a constant;

[0051] The activation function is calculated according to the following formula:

[0052] δ=ε·η n ;

[0053] Among them, δ represents the activation function and ε is a constant.

[0054] In some embodiments of the present invention, in the step of calculating the decision value corresponding to each decision solution based on the incentive function and the total time required for each decision solution, the decision value is calculated according to the following formula:

[0055] Δ=λ·δ-μ·T total ;

[0056] Where Δ represents the decision value, T total It represents the total time required for decision making, and λ and μ are both constants.

[0057] Another aspect of the present invention also provides a distributed reasoning device for low-orbit satellite constellations, which includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps implemented by the method described above.

[0058] Additional advantages, objects, and features of the present invention will be described in part in the following description and will become apparent to those skilled in the art after studying the following or may be learned by practice of the present invention. The objects and other advantages of the present invention may be particularly pointed out and attained in the description and drawings.

[0059] Those skilled in the art will understand that the purposes and advantages that can be achieved by the present invention are not limited to the above specific descriptions, and the above and other purposes that can be achieved by the present invention will be more clearly understood based on the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention.

[0061] Figure 1 A schematic diagram of an embodiment of the distributed reasoning method for low-orbit satellite constellations of the present invention;

[0062] Figure 2 It is a schematic diagram of the architecture of the prior art 1;

[0063] Figure 3 This is a schematic diagram of the architecture of prior art 2;

[0064] Figure 4 This is a schematic diagram of the architecture of this solution;

[0065] Figure 5 This is a schematic diagram of the implementation process of this plan. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0067] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, the accompanying drawings only show structures and / or processing steps closely related to the solutions according to the present invention, while other details that are not closely related to the present invention are omitted.

[0068] It should be emphasized that the term "include / comprises" when used herein refers to the existence of features, elements, steps or components, but does not exclude the existence or addition of one or more other features, elements, steps or components.

[0069] It should also be noted that, unless otherwise specified, the term "connection" herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0070] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0071] Introduction to existing technology:

[0072] To build a large-scale network of low-orbit satellite communication constellations, the number of low-orbit satellite launches will increase significantly. Currently, network technologies such as edge computing and computing-network fusion are gaining popularity. Integrating edge computing into satellite networks can significantly improve the quality of satellite communications. However, applying satellite networks to artificial intelligence (AI) scenarios, such as deep learning (DL) and deep neural network (DNN) reasoning, presents various challenges. For example, in large-scale LEO constellations, computing resources on satellites are limited due to various environmental factors in space. However, services such as DNN reasoning require significant computing and storage resources, making it difficult to perform DNN reasoning directly on the satellites. To address this issue, leveraging the principle of distributed reasoning in IoT networks to address the dynamic nature and resource constraints of IoT edge devices, this approach is introduced in large-scale LEO constellations, providing a solution for more efficient application in AI and other scenarios.

[0073] Prior Art 1

[0074] Cloud computing, as a further development of distributed computing and parallel processing within the internet environment, offers advantages such as resource scalability, service quantification, resource pooling, and transparency. Cloud computing addresses the storage and processing of massive amounts of situational data in satellite communication networks. It can perform comprehensive situation assessments and equipment failure rate statistics based on historical data. It can also extract symptom information based on observed events and then match it with knowledge bases to facilitate fault diagnosis and decision-making.

[0075] Satellite communication architecture integrated with cloud computing Figure 2As shown in the figure, Hadoop is a distributed computing framework that can be deployed on inexpensive hardware. It offers high fault tolerance and reliability, making it suitable for applications with large datasets. The underlying infrastructure comprises inexpensive physical machines, providing computing and storage capabilities for the system. Data agents monitor device parameters and link status information for satellite communication networks, ground networks, and optical networks. This information is collected and stored as metadata in the Hadoop cluster. Resources are then integrated and virtualized into a cloud computing resource pool, isolating software applications from the underlying devices, expanding hardware capacity, and simplifying software configuration. The computing and storage resources in this virtualized resource pool are used for cloud-based inference, enabling the application of satellite communications in artificial intelligence scenarios.

[0076] Disadvantages of prior art 1

[0077] Implementing DNN inference with the assistance of cloud computing can avoid the drawbacks of using mobile or embedded devices to perform local DNN inference directly, which lacks the support of large amounts of computing and storage resources. However, in the process of using cloud assistance, the performance of direct local DNN inference is abandoned, and the disadvantage of unpredictable remote server status is created. It also causes additional delays in remote data transmission with the ground cloud center. Moreover, private and sensitive information may be exposed during the ground data transmission and remote processing, which will lead to some deficiencies in data security. In addition, there are more and more application scenarios that require large-scale computing. In the future, situations awareness, holographic communication and other businesses will inevitably lead to the rapid explosion of generated and collected data to a scale that makes cloud-based centralized data processing difficult to achieve.

[0078] Prior Art 2

[0079] In recent years, Low Earth Orbit-Satellite Networks (LEO-SNs) have gradually taken a central role in satellite communications due to their short satellite-to-ground transmission distances and low construction costs. Satellite-based on-orbit computing can support latency-sensitive services by avoiding the significant overhead and latency of transmitting data to ground-based cloud centers. Offloading computing tasks directly to satellites can significantly reduce transmission latency. However, offloading services to a single LEO (Low Earth Orbit) satellite results in excessively high computing latency, which also fails to meet the low latency requirements of the services. Therefore, with the growing demand for computing services, it is necessary to utilize the computing resources of the entire LEO-SN for collaborative computing between multiple satellites, thereby simultaneously meeting both low transmission and computing latency requirements.

[0080] The basic architecture of satellite cloud network is as follows Figure 3As shown in the figure, the system mainly consists of a satellite fog layer and a ground cloud layer. The satellite fog layer includes a cluster of LEO satellites with abundant computing resources, which can provide sufficient computing services to ground users. The ground cloud layer is composed of various users, such as ocean users, desert users, and air users. These users are connected to the satellite fog layer through ground servers, enabling real-time control of satellites. Based on the current inter-satellite network status and service requirements, satellite computing resources are coordinated and scheduled to meet the resource needs of current services.

[0081] Disadvantages of Existing Technology 2

[0082] In satellite networks, due to the high-speed mobility of satellites, it is difficult to ensure that each satellite is always directly above the ground station it serves. However, the collaborative computing in the LEO-SN of the above-mentioned solution cannot support the continued service to the ground station when the communication signal between the satellite and the ground station is weakened due to satellite movement. This solution does not consider the service continuity of computational offloading and cannot solve the problem of how to complete information backhaul and processing in the inter-satellite link when the ground station is missing. Furthermore, due to the limited computing power and cache capacity per satellite, existing technologies are significantly limited. The adaptive segmentation of computing tasks, the optimal allocation of segmented tasks, the optimal storage of cache files, and the distributed storage of cache files in collaborative computing all consume a large amount of resources, requiring the design of a suitable solution to optimize the overall strategy.

[0083] This solution aims to address the above-mentioned huge challenges faced by the important development trend of applying artificial intelligence reasoning to giant low-orbit constellations, and conducts research on distributed reasoning methods for giant low-orbit constellations. However, the current research on giant low-orbit constellation networks is in its infancy, and there is no good solution for how to further expand the application of giant low-orbit constellations in artificial intelligence scenarios. The solution of the present invention focuses on the application scenarios of giant low-orbit constellation networks and explores methods for implementing distributed reasoning in this scenario. First, data is transmitted between satellites to avoid reduced data security caused by ground transmission. On the other hand, multiple satellites are enabled to complete tasks collaboratively, solving the problem of insufficient computing power of a single satellite and realizing the intelligent application of resources in giant low-orbit constellations.

[0084] To solve the above problems, Figure 1 As shown, the present invention proposes a distributed reasoning method for low-orbit satellite constellations, the method comprising the following steps:

[0085] Step S100: any satellite in the low-orbit satellite constellation receives a processing task, and the satellite that receives the processing task divides the processing task into a plurality of subtasks according to a preset splitting node;

[0086] In some embodiments of the present invention, the processing task is a preset task type, and any satellite in the low-orbit satellite constellation is provided with a processing task model. Based on the task type of the processing task, each of the processing tasks is preset with the same splitting node in any satellite, and the processing task is divided into multiple subtasks for solving different stages of the processing task based on the splitting node. If the task includes a first convolution layer and a second convolution layer executed sequentially, the calculations of the first convolution layer and the second convolution layer can be respectively used as subtasks.

[0087] In some embodiments of the present invention, a ground server receives a user request and sends a processing task corresponding to the user request to a low-orbit satellite constellation.

[0088] Step S200, constructing multiple decision-making schemes based on multiple subtasks and multiple satellites, wherein the combination of multiple subtasks allocated to multiple satellites in each decision-making scheme is different;

[0089] In some embodiments of the present invention, the low-orbit satellite constellation includes multiple small satellite clusters. The satellite that receives the processing task distributes multiple subtasks to satellites in the small satellite cluster. The satellite that processes the subtasks can process multiple subtasks simultaneously.

[0090] In the specific implementation process, the number of satellites in the giant low-orbit constellation is huge. Considering the objective environmental factors of geographical distance and resource limitation, the communication between satellites or between satellites and ground stations will be affected by the environment and cause transmission loss. In the giant low-orbit constellation distributed inference framework, a centralized and distributed hybrid structure is adopted. The specific system is as follows Figure 4 As shown in the figure, the entire giant low-orbit constellation is divided into multiple small satellite clusters. By dividing and offloading DNN tasks, the low-orbit satellites in the cluster perform DNN calculations in a distributed architecture and collaborate with ground servers to complete DNN tasks. Due to the influence of geographical distance, each low-orbit satellite only completes the DNN tasks assigned to its own cluster and tries to interact with its relatively closer neighboring low-orbit satellites to avoid problems such as low communication efficiency caused by long communication distances. In addition, due to limited satellite resources and the fixed location of ground servers, and the giant low-orbit constellation is always in high-speed motion, to ensure the effectiveness of task division and offloading, a ground server is set to complete collaborative task calculations only with the small satellite cluster closest to it.

[0091] In a specific implementation process, a plurality of decision-making schemes can be obtained by arranging and combining a plurality of subtasks and a plurality of satellites.

[0092] In a specific implementation, if multiple satellites include a1, a2, and a3, and multiple subtasks include b1, b2, and b3, then the various decision-making schemes include: 1. Allocate b1 to a1, b2 to a2, and b3 to a3; 2. Allocate b1 to a1, b2 to a3, and b3 to a2; 3. Allocate b1 to a3, b2 to a2, and b3 to a1; 4. Allocate b1 to a2, b2 to a1, and b3 to a3;

[0093] And execute the subtasks in the order of b1, b2 and b3.

[0094] Step S300, calculating the processing time required to complete each subtask based on the subtask packet size and the processing resource size of the satellite to which the subtask is assigned, calculating the transmission delay of the subtask between satellites based on the transmission speed between satellites, and calculating the total time required for each decision solution based on the processing time and transmission time;

[0095] Step S400, calculating a relative value function based on the sampling rate of the satellite receiving the processing task, and calculating an incentive function based on the relative value function;

[0096] During the specific implementation process, the sampling rate is adjusted according to the relative value function set by the low-orbit satellite, which affects the amount of resources required to process the input data packets. When a large amount of data is required, the sampling rate can be increased to ensure that there is sufficient data for deep neural network inference, thereby improving the activation function, but the delay in processing deep neural network requests will increase; when a large amount of data is not required, the sampling rate is reduced to reduce energy consumption and resource costs, and reduce processing delays, but it will affect the activation function.

[0097] Step S500 , calculating a decision value corresponding to each decision solution based on an incentive function and a total time required for each decision solution, and screening multiple decision solutions to obtain a final decision solution based on the decision value.

[0098] During the specific implementation process, the decision plan corresponding to the largest decision value is taken as the final decision plan.

[0099] Adopting the above scheme, this scheme divides the entire processing task into multiple subtasks, and corresponds to multiple decision-making schemes for the allocation of subtasks. In the process of selecting the decision-making scheme, it can comprehensively consider multiple indicators such as the effectiveness of satellite data collection, task transmission delay and processing delay. For task offloading and resource scheduling decisions, a quantum evolutionary algorithm based on incentive function and processing delay is designed, which can improve the processing efficiency of subtasks on the assigned satellites. Multiple satellites jointly solve the processing tasks, avoiding the problem of insufficient computing power of a single satellite, ensuring the processing effectiveness of processing tasks in low-orbit constellations and reducing complexity.

[0100] In some embodiments of the present invention, the processing resource size includes the number of CPU cores, and the step of calculating the processing time consumed to complete each subtask based on the data packet size of the subtask and the processing resource size of the satellite to which the subtask is assigned includes:

[0101] The number of CPU cores of the satellite to which the subtask is assigned is calculated based on the data packet size of the subtask and the number of CPU cores of the satellite to which the subtask is assigned;

[0102] The subtask processing rate is calculated based on the number of CPU cores of the satellite to which the subtask is assigned and the task processing rate of a single core in the satellite to which the subtask is assigned;

[0103] The processing time consumed by the satellite assigned the subtask to complete the subtask is calculated based on the subtask processing rate and the data packet size of the subtask.

[0104] In some embodiments of the present invention, in the step of calculating the number of CPU cores required to occupy the satellite to which the subtask is assigned based on the packet size of the subtask and the number of CPU cores of the satellite to which the subtask is assigned, the number of CPU cores required to occupy the satellite to which the subtask is assigned is calculated according to the following formula:

[0105]

[0106] Among them, l m,m (i) represents the number of CPU cores of satellite m required for subtask i to be assigned to satellite m by satellite n that receives the processing task, C n (i) represents the size of the subtask data packet received from satellite n, L m represents the total number of CPU cores of satellite m, ρ i Indicates the transmission parameters of subtask i, [] indicates rounding, a n,m (i) is the judgment parameter. If subtask i is assigned to satellite m by satellite n, then a n,m (i)=1, otherwise a n,m (i)=0.

[0107] In some embodiments of the present invention, [x] is a Gaussian function, which means an integer not exceeding (x+1).

[0108] In some embodiments of the present invention, the transmission parameter ρ of the subtask i is i It is a preset constant parameter.

[0109] In some embodiments of the present invention, in the step of calculating the subtask processing rate based on the number of CPU cores required to occupy the satellite to which the subtask is assigned and the task processing rate of a single core in the satellite to which the subtask is assigned, the subtask processing rate is calculated according to the following formula:

[0110] fn,m (i) = f m ·l n,m (i);

[0111] Among them, f n , m (i) represents the task processing rate of subtask i assigned by satellite n to satellite m on satellite m, f m represents the task processing rate of a single core in satellite m assigned subtask i, l n,m (i) represents the number of CPU cores of satellite m required for subtask i to be assigned to satellite m by satellite n that receives the processing task.

[0112] In some embodiments of the present invention, in the step of calculating the processing time consumed by the satellite assigned the subtask to complete the subtask based on the subtask processing rate and the data packet size of the subtask, the processing time consumed by the satellite assigned the subtask to complete the subtask is calculated according to the following formula:

[0113]

[0114] in, represents the processing time consumed by subtask i assigned by satellite n to satellite m on satellite m, C n (i) represents the size of the subtask data packet received from satellite n, σ represents the CPU size required by the CPU in satellite m to process each bit of subtask data, and f n,m (i) represents the task processing rate of subtask i assigned by satellite n to satellite m on satellite m, a n,m (i) is the judgment parameter. If subtask i is assigned to satellite m by satellite n, then a n,m (i)=1, otherwise a n,m (i)=0.

[0115] In some embodiments of the present invention, the CPU size σ required for the CPU in satellite m to process each bit of subtask data is a preset constant parameter.

[0116] In some embodiments of the present invention, the step of calculating the transmission delay of the subtask between satellites based on the transmission speed between satellites includes:

[0117] Calculate the transmission speed between satellites based on the transmission gain between satellites and the bandwidth of the inter-satellite link;

[0118] The transmission delay between satellites is calculated based on the transmission speed between satellites and the packet size of the subtask.

[0119] In some embodiments of the present application, in the step of calculating the transmission speed between satellites based on the transmission gain between satellites and the inter-satellite link bandwidth, the transmission speed between satellites is calculated according to the following formula:

[0120]

[0121] wherein, represents the transmission speed of data from satellite u to satellite m, W u,m represents the inter-satellite link bandwidth between satellite u and satellite m, P u represents the transmission power of satellite u, h u,m represents the transmission gain from satellite u to satellite m, N z represents a noise parameter.

[0122] In some embodiments of the present application, the noise parameter is a preset constant.

[0123] In some embodiments of the present application, the transmission gain between satellites is calculated according to the following formula:

[0124]

[0125] wherein, h u,m represents the transmission gain from satellite u to satellite m, represents the transmission gain of satellite u, represents the reception gain of satellite m, represents a free space path loss parameter between satellite u and satellite m;

[0126] In some embodiments of the present application, and are both preset constants.

[0127] The free space path loss parameter between satellite u and satellite m is calculated according to the following formula:

[0128]

[0129] wherein, represents the free space path loss parameter between satellite u and satellite m, d u,m represents the transmission distance between satellite u and satellite m, c represents the value of light speed, p u,m represents the communication frequency between satellite u and satellite m.

[0130] In some embodiments of the present application, in the step of calculating the transmission delay of transmission between satellites based on the transmission speed between satellites and the data packet size of sub-tasks, the transmission delay of transmission between satellites is calculated according to the following formula:

[0131]

[0132] in, represents the transmission delay between satellite u performing subtask k and satellite m performing subtask i, C n (i) represents the size of the data packet received from satellite n for subtask i, represents the transmission speed of data from satellite u to satellite m, a n,m (i) and a n,u (k) are all judgment parameters. If subtask i is assigned to satellite m by satellite n, then a n,m (i)=1, otherwise a n,m (i) = 0, if subtask k is assigned to satellite u by satellite n, then a n,u (k)=1, otherwise a n,u (k)=0.

[0133] During the specific implementation process, in the low-orbit satellite domain, a large number of satellites are distributed in low-Earth orbit according to the position planning at the time of launch. Satellites transmit data and signals through inter-satellite links. Currently, only communication between adjacent satellites is considered, and each satellite collects data at a different sampling rate for inference calculations.

[0134] In the specific implementation process, for two subtasks i and j with strong dependencies, i is the subtask executed in the previous step of j, then there is a relationship:

[0135] C n (j) = ρ i ·C n (i);

[0136] Among them C n (i) represents the packet size of subtask i from satellite n, C n (j) represents the packet size of subtask j from satellite n, ρ i is the transfer function of subtask i, and σ represents the CPU size required by the satellite CPU to process each bit of DNN data.

[0137] The transfer function may be a preset constant parameter.

[0138] In some embodiments of the present invention, the step of calculating the total time required for each decision solution based on the processing time and transmission time includes:

[0139] Calculate the end time parameter of each subtask in the decision solution based on the processing time and the transmission time;

[0140] The maximum end time parameter is used as the total time parameter required for the decision-making plan.

[0141] In some embodiments of the present invention, in the step of calculating the end time parameter of each subtask in the decision solution based on the processing time and the transmission time, the end time parameter of each subtask is calculated according to the following formula:

[0142]

[0143] in, represents the end time parameter of subtask i executed on satellite m, The end time parameter of subtask k executed on satellite u, represents the transmission delay between satellite u performing subtask k and satellite m performing subtask i, represents the processing time consumed by subtask i assigned by satellite n to satellite m on satellite m, and subtask k is the subtask executed before subtask i. If subtask i includes multiple previous subtasks, the maximum value in [] is taken.

[0144] In the specific implementation process, since the results of DNN inference are generally very small, the download delay of the final result can be excluded from the calculation of the overall delay.

[0145] like Figure 5 As shown, Figure 5 Where v2, v3, and v4 represent subtasks, and e represents transmission. In the specific implementation process, if the previous step of subtask v4 includes multiple subtasks v2 and v3, the time when the data is last transmitted to the satellite where v4 is located in the multiple previous subtasks is calculated as the start time of subtask v4.

[0146] By adopting the above scheme, the end time of each subtask can be calculated one by one in an iterative manner, and the end time of the subtask that ends last is the end time of the decision-making scheme.

[0147] In some embodiments of the present invention, in the steps of calculating the relative value function based on the sampling rate of the satellite receiving the processing task and calculating the incentive function based on the relative value function, the relative value function is calculated according to the following formula:

[0148] η n =α sp ;

[0149] Among them, η n represents the relative value function of the satellite n that receives the processing task, s sp represents the sampling rate of the satellite receiving the processing task, and α is a constant;

[0150] In the specific implementation process, considering that low-orbit satellites are distributed in different areas, a low-orbit constellation distributed reasoning domain is given, in which N low-orbit satellites are deployed, represented by the set N = {1,…n,..,N}, with different sampling rates {s1,s2,…,s n ,…,s N} (i.e. data sampling rate) collects data for inference calculation, thereby providing onboard inference services. Due to the different environments in which satellites are located, the effectiveness of the collected data will also vary. Therefore, η is set for satellites in different locations in the area. n Characterizing the relative value function of the data collected by satellite n, it can be obtained that the sampling rate is proportional to the relative value function.

[0151] In the satellite distributed reasoning domain, low-orbit satellites communicate with each other via inter-satellite links (ISLs). Since signal transmission loss in space is difficult to ignore, to ensure data integrity and the accuracy of DNN task reasoning, it is assumed that each satellite only offloads and schedules tasks with its neighboring satellites. That is, when the satellites in the cluster have sufficient computing resources to process DNN requests, satellite n can offload some of the DNN request computation to the collaborative satellite for task reasoning, thereby achieving distributed reasoning between satellites.

[0152] Furthermore, the sampling rate is used as a characteristic value of the request to measure the CPU size required to process the DNN request. Setting different sampling rates results in different CPU requirements, thus affecting the overall latency of the request processing. To ensure that tasks can be offloaded to collaborative satellites when local satellite computing resources are insufficient, tasks are split according to the strong dependencies between DNN layers, ensuring that the output of the previous layer serves as the input of the next layer in the DNN inference calculation.

[0153] In the specific implementation process, the above steps of calculating and obtaining the final decision solution are all performed on the satellite that receives the processing task.

[0154] The activation function is calculated according to the following formula:

[0155] δ=ε·η n ;

[0156] Among them, δ represents the activation function and ε is a constant.

[0157] In some embodiments of the present invention, in the step of calculating the decision value corresponding to each decision solution based on the incentive function and the total time required for each decision solution, the decision value is calculated according to the following formula:

[0158] Δ=λ·δ-μ·T total ;

[0159] Among them, Δ represents the decision value, δ represents the activation function, T total It represents the total time required for decision making, and λ and μ are both constants.

[0160] In the specific implementation process, after the satellite that receives the processing task completes the task offloading and computing, it transmits its inference calculation results to one or more collaborative satellites to continue to complete the next step of inference work. And so on. After completing the last step of inference calculation, the local satellite can directly download the inference results from the last collaborative satellite and use them to provide services.

[0161] When a collaborative satellite receives a DNN request from another satellite, it allocates resources to process it based on its own computational load. When the decision is made to offload tasks to a specific satellite, the DNN request and corresponding data to be processed on that satellite are already allocated. Resource scheduling is performed based on the size of the input data packets transmitted to the satellite, allocating appropriate resources for DNN inference calculations.

[0162] When multiple DNN requests arrive at the collaborative satellite at the same time, the multiple DNN requests will be processed in parallel, and resources will be allocated according to the proportion of each task scheduled to the satellite to the total task volume.

[0163] In the specific implementation process, when a local satellite collects a certain amount of data and generates a DNN request, it is assumed that all models used to infer the DNN request have been stored on all satellites, that is, all satellites are capable of processing all DNN requests to provide services. Due to the extremely short processing latency during DNN inference, it is assumed that the DNN task processing rate provided by the satellite's single-core CPU remains unchanged during the inference process.

[0164] To accelerate DNN inference, each DNN model is divided into several parts, one part of which is processed on the local satellite, and the other part is offloaded to the collaborative satellite of the local satellite for distributed inference. Sufficient resources are allocated to accommodate sufficient DNN requests based on the implementation resource status on the collaborative satellite.

[0165] In the specific implementation process, when solving the resource scheduling and task offloading problems of distributed DNN reasoning in low-orbit satellite networks, it is necessary to make the most favorable scheduling decisions possible based on the processing latency and incentive function of the DNN tasks to maximize the utility of the optimization goal. To solve the resource scheduling and task offloading problems of distributed DNN reasoning in low-orbit satellite networks, by simultaneously considering the incentive function and processing latency, a quantum evolutionary algorithm is used to iteratively determine the various parameters in the calculation process.

[0166] When running a quantum evolutionary algorithm based on an activation function and processing delay, the solution to the scheduling offloading decision may result in the same subtask being offloaded to different satellites multiple times or not being offloaded at all, thus failing to be processed. Therefore, by modifying the scheduling offloading decision steps, we can ensure that the solution obtained in each iteration satisfies the requirement that any subtask must be offloaded only once. By locally migrating the storage offloading decision, we can achieve the extension of the local optimal solution to the global optimal solution.

[0167] With the increasing computing power of low-orbit satellites (LEOs), existing technologies aim to use them to implement edge computing to meet the massive computing power demands of current services. Furthermore, with the increasing application of artificial intelligence (AI), the demand for intelligent services provided by satellites, such as satellite reconnaissance, is also growing. This makes the use of deep neural networks (DNNs) for task reasoning on satellites inevitable. LEO satellites can sense and collect data and promptly feed it back to the DNNs, which then process the data and output reasoning results. To address this issue, this solution leverages the principle of distributed reasoning in IoT networks to address the dynamic nature and resource constraints of IoT edge devices. We design a distributed reasoning system model, task offloading, and resource scheduling solution for large LEO constellations. This approach introduces distributed reasoning methods within large LEO constellations to address the perception of large-scale, distributed, diverse, complex, and noisy scenarios, enabling the large-scale data processing capabilities of LEO satellite constellations and providing a solution for their more efficient application in AI and other scenarios.

[0168] The beneficial effects of this program include:

[0169] 1. In response to the surge in demand for artificial intelligence services in low-orbit satellites, we designed a distributed DNN inference architecture for giant low-orbit constellations by combining deep neural networks with edge computing technologies. This architecture comprehensively considers the technical challenges faced in the context of low-orbit satellites and enables fine-grained partitioning of DNN models.

[0170] 2. Based on the architecture of DNN distributed inference in giant low-orbit constellations, and taking into account multiple indicators such as satellite data acquisition effectiveness, task transmission latency, and processing latency, a quantum evolutionary algorithm based on incentive functions and processing latency is designed for task offloading and resource scheduling decisions. This ensures the processing efficiency of DNN request tasks in low-orbit constellations and reduces complexity.

[0171] 3. This solution takes into account the overall network performance of low-orbit satellites, introduces an incentive function that is proportional to the relative value function of the sampling rate, and describes the problem as a joint optimization problem of the incentive function and processing delay, providing lower service delay and higher overall utility, thereby improving the efficiency of providing intelligent services in low-orbit satellite scenarios.

[0172] 4. To enable flexible and efficient resource sharing among multiple heterogeneous satellite nodes in a satellite networking environment, this solution's architecture enables resource scheduling and task offloading for low-orbit satellites through hierarchical and domain-specific controllers. Furthermore, this solution incorporates the data sampling rate of satellite nodes and establishes a relative value function based on this sampling rate to evaluate the benefits of distributed inference in the context of large low-orbit constellations.

[0173] 5. This solution considers the impact of the sampling rate of LEO satellite data on service latency. By introducing a relative value function and an incentive function to optimize the sampling rate, this approach achieves a dynamically adaptive LEO satellite data sampling rate. This solution, through multiple iterations of a quantum evolutionary algorithm, continuously approaches the optimal solution, solving the joint optimization problem of sampling rate and service latency, thereby improving the efficiency of providing intelligent services in LEO satellite scenarios.

[0174] An embodiment of the present invention also provides a distributed reasoning device for a low-orbit satellite constellation, which includes a computer device, the computer device including a processor and a memory, the memory storing computer instructions, and the processor being used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps implemented by the method described above.

[0175] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the computer program implements the steps implemented by the aforementioned distributed reasoning method for low-orbit satellite constellations. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0176] It should be understood by those skilled in the art that the various exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is specifically performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments that are used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0177] It should be understood that the present invention is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, a detailed description of known methods is omitted. In the above embodiments, several specific steps are described and illustrated as examples. However, the method of the present invention is not limited to the specific steps described and illustrated. Those skilled in the art may make various changes, modifications, and additions, or change the order of the steps after understanding the spirit of the present invention.

[0178] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or replace features of other embodiments.

[0179] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations to the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A distributed reasoning method for low-orbit satellite constellations, characterized in that: The steps of the method include: Any satellite in the low-orbit satellite constellation receives the processing task, and the satellite that receives the processing task divides the processing task into multiple subtasks according to preset splitting nodes; Construct multiple decision-making schemes based on multiple subtasks and multiple satellites, and each decision-making scheme has a different combination of multiple subtasks allocated to multiple satellites; The processing time consumed to complete each subtask is calculated based on the data packet size of the subtask and the processing resource size of the satellite to which the subtask is assigned, wherein the processing resource size includes the number of CPU cores. The number of CPU cores of the satellite to which the subtask is assigned is calculated based on the data packet size of the subtask and the number of CPU cores of the satellite to which the subtask is assigned; the subtask processing rate is calculated based on the number of CPU cores of the satellite to which the subtask is assigned and the task processing rate of a single core in the satellite to which the subtask is assigned; the processing time consumed by the satellite to which the subtask is assigned to complete the subtask is calculated based on the subtask processing rate and the data packet size of the subtask; the transmission delay of the subtask between satellites is calculated based on the transmission speed between satellites, and the total time required for each decision plan is calculated based on the processing time and the transmission time; Calculate a relative value function based on a sampling rate of a satellite receiving a processing task, and calculate an incentive function based on the relative value function; A decision value corresponding to each decision solution is calculated based on the incentive function and the total time required for each decision solution, and a final decision solution is obtained by screening multiple decision solutions based on the decision value.

2. The distributed reasoning method for low-orbit satellite constellations according to claim 1, characterized in that: In the step of calculating the number of CPU cores of the satellite to which the subtask is assigned based on the data packet size of the subtask and the number of CPU cores of the satellite to which the subtask is assigned, the number of CPU cores of the satellite to which the subtask is assigned is calculated according to the following formula: Among them, l n,m (i) represents the number of CPU cores of satellite m required for subtask i to be assigned to satellite m by satellite n that receives the processing task, C n (i) represents the size of the subtask data packet received from satellite n, L m represents the total number of CPU cores of satellite m, ρ i Indicates the transmission parameters of subtask i, [] indicates rounding, a n,m (i) is the judgment parameter. If subtask i is assigned to satellite m by satellite n, then a n,m (i)=1, otherwise a n,m (i) = 0; In the step of calculating the subtask processing rate based on the number of CPU cores of the satellite to which the subtask is assigned and the task processing rate of a single core in the satellite to which the subtask is assigned, the subtask processing rate is calculated according to the following formula: f n,m (i)=f m ·l n,m (i); Among them, f n,m (i) represents the task processing rate of subtask i assigned by satellite n to satellite m on satellite m, f m represents the task processing rate of a single core in satellite m that is assigned subtask i.

3. The distributed reasoning method for low-orbit satellite constellations according to claim 1, characterized in that: In the step of calculating the processing time consumed by the satellite assigned the subtask to complete the subtask based on the subtask processing rate and the data packet size of the subtask, the processing time consumed by the satellite assigned the subtask to complete the subtask is calculated according to the following formula: in, represents the processing time consumed by subtask i assigned by satellite n to satellite m on satellite m, C n (i) represents the size of the subtask data packet received from satellite n, σ represents the CPU size required by the CPU in satellite m to process each bit of subtask data, and f n,m (i) represents the task processing rate of subtask i assigned by satellite n to satellite m on satellite m, a n,m (i) is the judgment parameter. If subtask i is assigned to satellite m by satellite n, then a n,m (i)=1, otherwise a n,m (i)=0.

4. The distributed reasoning method for low-orbit satellite constellations according to claim 1, characterized in that: The step of calculating the transmission delay of the subtask between satellites based on the transmission speed between satellites includes: Calculate the transmission speed between satellites based on the transmission gain between satellites and the bandwidth of the inter-satellite link; The transmission delay between satellites is calculated based on the transmission speed between satellites and the packet size of the subtask.

5. The distributed reasoning method for low-orbit satellite constellations according to claim 4, characterized in that: In the step of calculating the transmission speed between satellites based on the transmission gain between satellites and the inter-satellite link bandwidth, the transmission speed between satellites is calculated according to the following formula: in, W represents the transmission speed of data from satellite u to satellite m. u,m represents the intersatellite link bandwidth between satellite u and satellite m, P u represents the transmission power of satellite u, h u,m represents the transmission gain from satellite u to satellite m, N z represents the noise parameter; In the step of calculating the transmission delay between satellites based on the transmission speed between satellites and the data packet size of the subtask, the transmission delay between satellites is calculated according to the following formula: in, represents the transmission delay between satellite u performing subtask k and satellite m performing subtask i, C n (i) represents the size of the data packet received from satellite n for subtask i, a n,m (i) and a n,u (k) are all judgment parameters. If subtask i is assigned to satellite m by satellite n, then a n,m (i)=1, otherwise a n,m (i) = 0, if subtask k is assigned to satellite u by satellite n, then a n,u (k)=1, otherwise a n,u (k)=0.

6. The distributed reasoning method for low-orbit satellite constellations according to claim 1, characterized in that: The step of calculating the total time required for each decision solution based on the processing time and transmission time includes: Calculate the end time parameter of each subtask in the decision solution based on the processing time and the transmission time; The maximum end time parameter is used as the total time parameter required for the decision-making plan.

7. The distributed reasoning method for low-orbit satellite constellations according to claim 6, characterized in that: In the step of calculating the end time parameter of each subtask in the decision solution based on the processing time and the transmission time, the end time parameter of each subtask is calculated according to the following formula: in, represents the end time parameter of subtask i executed on satellite m, The end time parameter of subtask k executed on satellite u, represents the transmission delay between satellite u performing subtask k and satellite m performing subtask i, represents the processing time consumed by subtask i assigned by satellite n to satellite m on satellite m, and subtask k is the subtask executed before subtask i. If subtask i includes multiple previous subtasks, the maximum value in [] is taken.

8. The distributed reasoning method for low-orbit satellite constellations according to claim 1, characterized in that: In the steps of calculating the relative value function based on the sampling rate of the satellite receiving the processing task and calculating the incentive function based on the relative value function, the relative value function is calculated according to the following formula: or n =αs sp ; Among them, η n represents the relative value function of the satellite n that receives the processing task, s sp represents the sampling rate of the satellite receiving the processing task, and α is a constant; The activation function is calculated according to the following formula: d=e·h n ; Among them, δ represents the activation function and ε is a constant; In the step of calculating the decision value corresponding to each decision solution based on the incentive function and the total time required for each decision solution, the decision value is calculated according to the following formula: Δ=λ·δ-μ·T total ; Among them, Δ represents the decision value, δ represents the activation function, T total It represents the total time required for decision making, and λ and μ are both constants.

9. A distributed reasoning device for low-orbit satellite constellations, characterized in that: The apparatus includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, the processor is used to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the apparatus implements the steps implemented by the method according to any one of claims 1 to 8.

Images