Data intelligent calculation method for 6G cloud-edge-end edge intelligent satellite internet combined with low-orbit communication satellite

Abstract

The invention provides a data intelligent calculation method for a 6G cloud-edge-end edge intelligent satellite internet in combination with a low-orbit communication satellite, and belongs to the field of 6G satellite internet. The method comprises the following steps: deploying a laser emission device in a'air-space-ground-sea 'converged communication network, establishing three laser inter-satellite links of GEO-GEO, GEO-LEO and LEO-LEO, processing data by adopting a laser communication ranging integrated structure, and performing staged fine adjustment by combining a Markov transfer matrix method with a large model to predict and quantify the emission power of each network management and control instruction; and meanwhile, a cloud-edge-end edge intelligent framework is constructed, massive satellite data form grids for storage by an optimal model segmentation method, and the grids are distributed to the edge or the cloud for operation according to needs. The problems that satellite communication power management and control are difficult, regional coverage is limited, and mass data processing resources are insufficient are solved, and 6G satellite internet communication stability and data processing efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of 6G satellite internet technology, and in particular to a method for intelligent computing of low-orbit communication satellite data in conjunction with 6G cloud-edge-device edge intelligent satellite internet. Background Technology

[0002] With the continuous development of communication technology, 6G, as the next-generation communication technology after fifth-generation mobile communication technology, is currently in the research and development stage. 6G aims to solve the problem of limited coverage in sea, land, air, and space, expanding the breadth and depth of network communication technology in the human living environment, and promoting the extension of internet technology towards integrated air, land, and sea coverage. Building a satellite internet information network system is the core vision of the 6G era. This system mainly consists of interconnected satellite internet, terrestrial internet, and mobile communication networks. The ultimate goal is to build a network communication system with "global coverage, ubiquitous access, on-demand service, and security and reliability," achieving seamless global coverage and forming a comprehensive interconnected internet communication network for people, things, and events.

[0003] Cloud computing, with its advantages of high data transmission rates, strong information processing capabilities, low operating costs, and high flexibility, can effectively solve the storage and maintenance problems of traditional data processing systems, meet the stringent computing power requirements of space exploration activities, and build a bridge for long-distance data transmission. However, current satellite resources in the air are limited, and a large amount of learning work needs to be completed through cloud or ground computing. This not only leads to low data processing timeliness but also increases the complexity of cloud computing. At the same time, during satellite data processing, satellite receivers face the challenge of processing massive amounts of data, and limited computing and communication hardware resources, as well as data resources, make it difficult to support complex neural network training tasks. These problems have all constrained the development and application of 6G satellite internet technology. Summary of the Invention

[0004] Purpose of the invention: To propose a method for intelligent computing of low-orbit communication satellite data in 6G cloud-edge-device edge intelligent satellite internet, in order to solve the technical problems in existing 6G satellite internet such as unreasonable satellite communication power control, limited geographical coverage, low efficiency of massive data processing, and limited hardware and data resources that make it difficult to support complex neural network training, thereby improving the communication performance and data processing capabilities of 6G satellite internet.

[0005] This invention proposes a method for intelligent computing using low-Earth orbit communication satellite data combined with 6G cloud-edge-device edge intelligent satellite internet, the steps of which are as follows:

[0006] S1. Deploy laser emission devices in the "air-space-ground-sea" integrated communication network composed of satellite internet and constellations, and use AI analysis to predict the power trend of control commands;

[0007] S2. Construct a cloud-edge-device edge intelligence framework. Use the optimal model segmentation method to divide the massive satellite data from the receiving end into a grid of control commands, which contains multiple sub-grids. A hierarchical model aggregation strategy is used between the grids.

[0008] In a further embodiment, step S1 specifically includes:

[0009] S1-1. To address the different laser communication needs between satellite internet, constellations, and opposite terminals, three types of inter-satellite laser links are established to support laser transmitting devices, and an integrated laser communication and ranging structure is adopted to process and control laser data.

[0010] S1-2. Laser emission devices are deployed in each part of the "air-space-terrestrial-sea" integrated communication network, including the space-based satellite communication network, the air-based satellite communication network, the ground-based communication network, and the deep-sea ocean communication network.

[0011] S1-3. By using AI analysis to predict the power trend of control commands, the required transmission power of each network can be quantified.

[0012] In a further embodiment, the laser inter-satellite links in the three cases of step S1-1 are specifically as follows:

[0013] Laser communication between the ground terminal and the geostationary orbit platform (GEO-GEO);

[0014] GEO-LEO laser communication between the ground terminal and the geostationary orbit platform and the low Earth orbit platform;

[0015] Laser communication between the ground terminal and the low-orbit platform (LEO-LEO).

[0016] In a further embodiment, the integrated laser communication ranging structure in step S1-1 consists of a transmitting subsystem, a receiving subsystem, an optical antenna, an ATP subsystem, and a data processing and control subsystem. The system loads the information to be transmitted onto the modulator's exciter via a modulator. The exciter current of the modulator changes according to the signal variation. After the laser's output signal is modulated by the modulator, the relevant parameters change according to the corresponding rules. The optical antenna transmits the laser's output signal, the detector detects the laser signal, and the demodulator recovers the original information, thus completing the data transmission.

[0017] In a further embodiment, in steps S1-3, the power trend of the control command is predicted by AI analysis. Specifically, the Markov transition matrix method is used to predict the power consumption trend of the next control command of the same type based on the power required by the current control command. The formula of the Markov transition matrix method model is: X(k+1)=X(k)×P, where X(k) represents the state vector of the trend analysis and prediction object at time t=k, P represents the one-step transition probability matrix, and X(k+1) represents the state vector of the trend analysis and prediction object at time t=k+1.

[0018] In a further embodiment, steps S1-3 also involve fine-tuning a large model and performing differential calibration using Markov predictions. The fine-tuning of the large model specifically includes:

[0019] The first stage of fine-tuning involves analyzing the data from control commands reaching the ground terminal. Ridge regression overfitting loss is used to calculate the power required to mitigate the loss, providing a preliminary reference for power adjustments at the transmitter. The ridge regression overfitting loss calculation formula to prevent overfitting is as follows:

[0020]

[0021] Where L(θ) is the loss function; y i The actual value of the control command power; X i The input parameter matrix includes communication capacity and communication distance; θ is the fitting hyperparameter; α is the weight of the identity matrix, ranging from 0.01 to 0.1, adjusted according to data fluctuations; ||θ|| 2 This is an L2 regularization term used to prevent overfitting;

[0022] Based on the calculated L(θ), when L(θ) > the preset threshold, the required additional transmission power is calculated according to the power compensation value = L(θ) × the reference power coefficient, which serves as the initial basis for the adjustment of the transmitter.

[0023] The second-stage fine-tuning involves using the dataset composed of log data queries after ridge regression fitting and filtering as the computational task, and then calculating the perplexity level (PPL) after computational power optimization.

[0024]

[0025] Where N is the length of the text sequence, w i It is the i-th word in the sequence, p(w i |w 1:i-1 ) represents the probability of the i-th word given the preceding words.

[0026] In a further embodiment, the large model fine-tuning can be extended to three-stage, four-stage, or other multi-stage fine-tuning as needed; if the multi-stage PPL fluctuation is less than 10%, a weighted average is used to set the final PPL; if the fluctuation is greater than 10%, the adjustment result of the previous stage before the large fluctuation is retained.

[0027] In a further embodiment, the optimal model segmentation method in step S2 specifically includes:

[0028] The data usage density of each control subgrid is analyzed using data segmented by edge intelligent meshing. The network usage density is calculated using dynamically changing control command volume, data traffic, and connection count. The calculation formula is as follows:

[0029] Composite Density = w u User Density+w t Traffic Density+w c ConnectionDensity

[0030] Wherein, Composite Density is the density of the integrated control and command sub-network, User Density is the density of the number of control and command grids, Traffic Density is the density of sub-grids, Connection Density is the density of sub-grid connections, and w u w t and w c These are the weights used for the number of subgrids, control commands, and connections, respectively.

[0031] Based on the density of the sub-grid, adjust the grid size of each sub-grid, dividing high-density areas into larger grids and low-density areas into smaller grids. The calculation formula is as follows:

[0032]

[0033] Wherein, New Grid Size is the adjusted grid size, Base Grid Size is the initial grid size, Network Density is the network usage density of each subgrid, and Average Density is the average network usage density of the entire grid.

[0034] In a further embodiment, after constructing the cloud-edge-device edge intelligence framework in step S2, when the data related to the same type of control instructions in the massive satellite data is stored in the cloud, the computing pressure on the cloud platform is first alleviated by a grid composed of temporary storage networks of satellite edge nodes; the data related to the same type of control instructions form a large grid containing multiple sub-grids; if it is new data, it is transferred from the temporary storage point to the cloud for storage and computation, otherwise it remains in the temporary storage point.

[0035] In a further embodiment, the parameters of each NTN node in the "air-space-ground-sea" converged communication network are as follows:

[0036] GEO satellite: altitude 35,786 km. The power trend predicted by the power-Markov chain model is 30% increase and 70% decrease. The power trend predicted by the power-large model with fine-tuning test is 30% increase and 70% decrease with an accuracy of 90%. Expected use cases include wide coverage, media broadcasting, public content broadcasting, fixed / mobile cell backhaul connection, and communication for users in urban / remote areas.

[0037] LEO satellites: altitude 400-1600km, expected use cases include fixed / mobile cell backhaul connectivity, urban / remote area user communication, media multicast services, and wide area IoT services;

[0038] VLEO satellites: altitude 100-400km, expected use cases include mobile cell backhaul connectivity, communication for users in urban / remote areas, media multicast services, wide area IoT, and broadband internet;

[0039] HAPS: Altitude 15-25km, expected use cases include air / ground base station backhaul connection, urban / remote area user communication, media multicast service, and IoT application.

[0040] UAV: Altitude 0.1-10km, expected use cases include ground base station backhaul connection, air relay / TRP, hotspot video-on-demand, and regional emergency services.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) Optimize power control and improve communication stability: By establishing three types of laser inter-satellite links and adopting an integrated laser communication ranging structure, the stability and accuracy of laser data transmission are ensured. At the same time, by using the Markov transfer matrix method combined with the AI ​​prediction method of large model fine-tuning, the power trend of control commands is accurately predicted, the required transmission power of each network is quantified, and communication quality problems caused by excessive or insufficient satellite communication power are avoided, thereby improving the communication stability and reliability of the "air-space-ground-sea" integrated communication network.

[0043] (2) Achieve full coverage and expand application scenarios: Laser transmitting devices are deployed in the space-based, air-based, ground-based and deep-sea parts of the "air-space-ground-sea" integrated communication network. Combined with the characteristics of different NTN nodes, full communication coverage of sea, land, air and space is achieved, which meets the needs of various application scenarios such as wide coverage (eMBB, mMTC), media broadcasting, public content broadcasting, cell backhaul connection, communication for users in urban / remote areas, wide area Internet of Things, and regional emergency services.

[0044] (3) Efficiently process massive amounts of data and alleviate resource pressure: Construct a cloud-edge-device edge intelligence framework, adopt the optimal model segmentation method to perform gridded processing on massive satellite data, and adjust the grid size according to the data usage density to achieve reasonable data storage and allocation. At the same time, by combining temporary storage at edge nodes and cloud storage, the computing pressure on the cloud platform is effectively alleviated, and the dual constraints of hardware and data resources on massive data at the satellite receiver end are solved, enabling complex neural network training tasks to be carried out smoothly under limited resources, thereby improving data processing efficiency and overall network performance. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of an inter-satellite laser link.

[0046] Figure 2 This is a schematic diagram illustrating the working principle of integrated laser communication and ranging.

[0047] Figure 3 This is a schematic diagram of an integrated communication network encompassing air, space, land, and sea. Detailed Implementation

[0048] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0049] Example 1

[0050] This embodiment discloses a method for intelligent computing using 6G cloud-edge-device edge smart satellite internet combined with low-Earth orbit communication satellite data, including the following steps:

[0051] I. Deploy laser transmitting devices in the "air-space-terrestrial-sea" converged communication network composed of satellite internet and constellations, and use AI analysis to predict the power trend of control commands.

[0052] (1) Establishing three types of laser inter-satellite links to support the laser transmitting device: To address the different laser communication needs between satellite internet, constellations, and the counterpart terminal, three types of laser inter-satellite links are established: one between the ground terminal and the geostationary orbit platform (GEO-GEO), another between the ground terminal and both the geostationary orbit platform and the low Earth orbit platform (GEO-LEO), and a third between the ground terminal and the low Earth orbit platform (LEO-LEO), to support the stable operation of the laser transmitting device. Simultaneously, an integrated laser communication and ranging structure is adopted for laser data processing and control. This structure consists of a transmitting subsystem, a receiving subsystem, an optical antenna, an ATP (acquisition, tracking, and aiming) subsystem, and a data processing and control subsystem. During data transmission, the system loads the information to be transmitted onto the modulator's exciter via a modulator. The modulator's exciter current changes according to the signal variation. After the laser output signal is modulated by the modulator, its amplitude, phase, frequency, intensity, and other parameters change according to corresponding rules. The optical antenna transmits the laser output signal, the detector detects the laser signal, and the demodulator recovers the original information, completing the data transmission.

[0053] (2) Deployment of laser transmitting devices across the entire network: Laser transmitting devices are deployed in all components of the "air-space-terrestrial-sea" integrated communication network, namely the space-based satellite communication network, the air-based satellite communication network, the terrestrial communication network, and the deep-sea ocean communication network. Among them, the space-based satellite communication network is centered on satellite communication nodes and uses artificial Earth satellites as relay stations to realize wireless communication between different locations on Earth; the air-based satellite communication network is mainly composed of high-altitude communication platforms (HAPs) flying at an altitude of about 20km above the ground and large unmanned aerial vehicles; the terrestrial communication network consists of the terrestrial Internet and mobile communication networks, and is the basic network for providing communication services to a large number of customers under the "air-space-terrestrial integrated" network structure; the deep-sea ocean communication network consists of a maritime wireless communication system, a marine satellite communication system, and a shore-based mobile communication system based on a terrestrial cellular network, which together realize communication coverage of near-shore, offshore, and deep-sea areas.

[0054] (3) AI predicts the power trend of control commands: The Markov transition matrix method is used to predict the power trend of control commands. This method predicts the power consumption trend of the next control command of the same type based on the power required by the current control command of the transmitter. The model formula is X(k+1)=X(k)×P, where X(k) represents the state vector of the trend analysis and prediction object at time t=k, P represents the one-step transition probability matrix, and X(k+1) represents the state vector of the trend analysis and prediction object at time t=k+1. For example, if the initial transmit power trend state vector is [0.3, 0.7] (i.e., 30% increase, 70% decrease), the state vector with a 60% probability of increasing and a 40% probability of decreasing transmit power in this current trend is [0.6, 0.4], and the state vector with a 30% probability of increasing and a 70% probability of decreasing transmit power in this current trend is [0.3, 0.7]. Then, the predicted probability of the next transmit power trend increasing is 0.3 × 0.6 + 0.3 × 0.7 = 39%, and the predicted probability of the next transmit power trend decreasing is 0.3 × 0.4 + 0.7 × 0.7 = 61%.

[0055] Meanwhile, to improve prediction accuracy, a large model is fine-tuned and differentiated from Markov predictions. The large model fine-tuning process is divided into two stages.

[0056] During the first phase of fine-tuning, the data from control commands reaching the ground terminal is analyzed. Ridge regression overfitting loss is used to calculate the power required to mitigate the loss, providing a preliminary reference for power adjustment at the transmitter. The ridge regression overfitting loss calculation formula is as follows:

[0057]

[0058] Where L(θ) is the loss function; y i The actual value of the control command power; X i The input parameter matrix includes communication capacity and communication distance; θ is the fitting hyperparameter; α is the weight of the identity matrix, ranging from 0.01 to 0.1, adjusted according to data fluctuations; ||θ|| 2 This is an L2 regularization term used to prevent overfitting.

[0059] During the second-stage fine-tuning, the dataset consisting of the log data query set (containing information such as communication capacity, communication distance, bit error rate, and power loss due to atmospheric attenuation) filtered by ridge regression fitting is used as a computing power task. After computing power optimization, perplexity (PPL) is calculated. The perplexity calculation formula is as follows: Where N is the length of the text sequence, w i It is the i-th word in the sequence, p(w i |w 1:i-1) represents the probability of the i-th word given the preceding words. Depending on actual needs, the large model fine-tuning can be extended to three-stage, four-stage, or other multi-stage fine-tuning. If the multi-stage PPL fluctuation is less than 10%, a weighted average is used to set the final PPL; if the fluctuation is greater than 10%, the adjustment result of the stage before the largest fluctuation is retained. AI prediction quantifies the required emission power of each network, ensuring the laser emission device operates at a reasonable power.

[0060] II. Constructing a cloud-edge-device edge intelligence framework to process massive amounts of satellite data

[0061] A cloud-edge-device edge intelligence framework is constructed, which uses the optimal model segmentation method to form a grid of control commands from the massive satellite data received at the receiving end, and contains multiple sub-grids. A hierarchical model aggregation strategy is adopted between the grids.

[0062] (1) Implementation of the optimal model segmentation method: The edge intelligent temporary storage point consists of countless grid nodes. Based on the satellite data associated with the satellite receiver's control commands (including communication capacity, communication distance, bit error rate, and power loss due to atmospheric attenuation, etc.), the data is stored according to its respective grid. First, the data usage density of each control sub-grid is analyzed using the data after edge intelligent grid segmentation. The network usage density is calculated by dynamically changing the number of control commands, data traffic, and connection count. The calculation formula is as follows:

[0063] Composite Density = w u User Density+w t Traffic Density+w c ConnectionDensity

[0064] Where CompositeDensity is the density of the integrated control and command subnetwork, UserDensity is the density of the number of control and command grids, TrafficDensity is the density of subgrids, ConnectionDensity is the density of subgrid connections, and w u w t and w c These are weights used for the number of sub-grids, control commands, and connections, respectively, and these weights can be set according to actual conditions. Then, the grid size of each sub-grid is adjusted based on the sub-network density, dividing high-density areas into larger grids and low-density areas into smaller grids. This increases the proportion of grids with weaker signals and poorer control perception in the sub-grids, ensuring more accurate and meaningful network performance data. The calculation formula is as follows:

[0065]

[0066] Where NewGridSize is the adjusted grid size, BaseGridSize is the initial grid size, NetworkDensity is the network usage density of each subgrid, and AverageDensity is the average network usage density of the entire grid.

[0067] (2) Data Storage and Computation Allocation: In the cloud-edge-device edge intelligence framework, when similar control command-related data from massive satellite data is stored in the cloud, it is first stored in a grid composed of temporary storage networks of satellite edge nodes to alleviate the computing pressure on the cloud platform. Similar control command-related data form a large grid, which contains multiple sub-grids. If the data is new, it is transferred from the temporary storage point to the cloud for storage and computation; if the data is not new, it remains in the temporary storage point. In this way, the dual limitations of hardware and data resources on massive data at the satellite receiver are effectively solved, while addressing the challenge of limited computing and communication hardware resources and data resources making it difficult to support complex neural network training tasks.

[0068] Example 2

[0069] Based on Example 1, this example discloses some details.

[0070] Figure 1 The exhibition showcases the connection relationships of three types of laser inter-satellite links: GEO-GEO, GEO-LEO, and LEO-LEO, including GEO satellites (GEO1, GEO2), LEO satellites (LEO1, LEO2), and ground terminals. It clearly presents the laser communication links between different satellites and ground terminals, and intuitively reflects the construction method of laser inter-satellite links.

[0071] Figure 2 The structural composition of both the missile-borne laser terminal and the ground-based laser terminal is shown, including the transmitting subsystem, optical system, receiving subsystem, data processing and ATP subsystem, and control subsystem. The transmission paths of the communication and ranging signals are also labeled, along with processing steps such as codeword fitting and mL, providing a detailed explanation of the working principle of the integrated laser communication and ranging structure.

[0072] Figure 3 It presents the overall architecture of space-based satellite communication network, air-based communication network, ground-based communication network and deep-sea ocean communication network. The various network components are interconnected through laser data transmission, demonstrating the full picture of the "air-space-ground-sea" integrated communication network and reflecting the characteristics of full network coverage.

[0073] Deployment and power prediction of laser emission device in "air-space-ground-sea" integrated communication network

[0074] (1) Integrated Implementation of Laser Inter-Satellite Link Establishment and Laser Communication Ranging: In practical applications, based on the communication requirements of satellite internet, constellations, and ground terminals, three types of laser inter-satellite links are established: GEO-GEO, GEO-LEO, and LEO-LEO. Taking the GEO-GEO link as an example, a suitable GEO satellite (altitude 35786km) and ground terminal are selected, and laser transmitting and receiving devices are configured to ensure stable link connection. In the integrated laser communication ranging structure, the transmitting subsystem uses a high-performance laser, the receiving subsystem uses a high-sensitivity detector, and the optical antenna is selected with appropriate aperture and gain according to the communication distance and environmental conditions. The ATP subsystem ensures accurate transmission of laser signals through high-precision tracking and aiming technology. In data transmission testing, the test information is loaded onto the exciter via a modulator, and the exciter current is adjusted to simulate different signal changes. The laser output signal is modulated and transmitted by the optical antenna. The detector at the ground terminal receives the signal and recovers the information through a demodulator, testing the accuracy and rate of data transmission to ensure that communication requirements are met.

[0075] (2) Deployment of laser emitting devices across the entire network: In the space-based satellite communication network, suitable laser emitting devices are installed for each of the GEO, LEO, and VLEO satellites in orbit. Miniaturized and low-power devices are selected based on the satellite's payload, power consumption, and communication requirements. In the air-based satellite communication network, laser emitting devices are deployed on HAPs (altitude 15-25km) and large UAVs. Considering the special characteristics of the high-altitude environment, the devices must have the ability to resist low temperatures, radiation, and interference. In the ground-based communication network, laser emitting devices are integrated into ground base stations to achieve interconnection with the existing terrestrial Internet and mobile communication networks. In the deep-sea and ocean-going communication network, shipborne laser terminals are installed on ships to cooperate with marine satellite communication systems and shore-based mobile communication systems to achieve communication between ships and the coast, and between ships.

[0076] (3) AI power prediction implementation: Collect historical control command power data, including power consumption under different NTN nodes and different communication scenarios. Use the Markov transition matrix method to construct a power prediction model and determine the state vector and transition probability matrix. For example, according to historical data statistics, the initial transmission power trend is 30% increase and 70% decrease, that is, state vector X(0)=[0.3,0.7]; by analyzing the power change pattern in historical data, determine the one-step transition probability matrix P. Assuming that when the power increases, the probability of the next increase is 60% and the probability of the next decrease is 40%, and when the power decreases, the probability of the next increase is 30% and the probability of the next decrease is 70%, then the transition probability matrix P=[[0.6,0.4],[0.3,0.7]]. Using this model to predict the next power trend, we calculate X(1)=X(0)×P=[0.3×0.6+0.7×0.3,0.3×0.4+0.7×0.7]=[0.39,0.61], which means that the probability of the next power increasing is 39% and the probability of decreasing is 61%.

[0077] Simultaneously, a large model is used for fine-tuning. A suitable large model for communication data processing (such as a model based on the Transformer architecture) is selected, and historical control commands arriving at ground terminals (including information such as communication capacity, communication distance, bit error rate, and power loss due to atmospheric attenuation) is used as training data. In the first stage of fine-tuning, the ridge regression algorithm is used to fit the data, an appropriate regularization parameter 'a' is set, and the overfitting loss is calculated to quantify the power required to compensate for the lost portion. For example, by performing ridge regression fitting on a batch of data with a communication capacity of 10Gbps, a communication distance of 5000km, a bit error rate of 10^-6, and a power loss of 2dB due to atmospheric attenuation, the loss value is obtained, thus determining that 5W of additional power is needed, providing a reference for adjusting the power at the transmitting end. In the second stage of fine-tuning, the log data filtered after ridge regression fitting is organized into a data query set, which is used as input data for the large model. After computational quantification, the PPL value is calculated. If the PPL value is high, it indicates that the model's understanding and prediction ability of the data is poor, requiring adjustment of model parameters or addition of training data, and further fine-tuning until the PPL value reaches the expected target. If higher prediction accuracy is required in practice, a three-stage fine-tuning process can be performed. Based on the two-stage fine-tuning, more data from complex scenarios are added to further optimize the model. Finally, the final model is determined based on the fluctuation of PPL in the multi-stage process.

[0078] Cloud-Edge-Device Edge Intelligence Framework Construction and Massive Data Processing

[0079] (1) Implementation of the optimal model segmentation method: At the edge intelligent temporary storage point, satellite data associated with the satellite receiver control command is used, such as communication capacity (e.g., 2Gbps, 5Gbps, 10Gbps), communication distance (e.g., 1000km, 3000km, 5000km), and bit error rate (e.g., 10). -5 10 -6 10 -7 To mitigate the effects of atmospheric attenuation and power loss (e.g., 1dB, 2dB, 3dB), the data is divided into different grids. Weights w are set. u =0.3, w t =0.4, w c =0.3, calculate the composite density (CompositeDensity) for each subgrid. For example, if a subgrid has UserDensity = 5 (indicating 5 control / command grids), TrafficDensity = 8 (indicating a subgrid density of 8), and ConnectionDensity = 6 (indicating a subgrid connection density of 6), then CompositeDensity = 0.3 × 5 + 0.4 × 8 + 0.3 × 6 = 1.5 + 3.2 + 1.8 = 6.5. Assuming the average grid density of the entire grid is AverageDensity = 5, and the initial grid size BaseGridSize = 100m × 100m, the grid size of the high-density subgrid is calculated as NewGridSize = 100 × (1 + (6.5 - 5) / 5) = 100 × (1 + 0.3) = 130m × 130m. The grid size of this high-density subgrid is then adjusted to 130m × 130m. For low-density subgrids, such as CompositeDensity = 3, the grid size is calculated as NewGridSize = 100 × (1 + (3 - 5) / 5) = 100 × (1 - 0.4) = 60m × 60m. The grid size is then reduced.

[0080] (2) Data Storage and Computation Allocation: In actual operation, after receiving massive amounts of data, the satellite receiver first classifies the data and sends data related to similar control instructions to the temporary storage network of the edge nodes. The edge nodes allocate the data to the corresponding subgrids for storage according to the optimal model partitioning method. For example, control instruction data with a communication capacity of 5Gbps and a communication distance of 3000km are allocated to the corresponding subgrids. The data in the temporary storage is periodically checked to determine whether it is new data. If data is detected that does not exist in the historical database, it is marked as new data and transmitted to the cloud for storage via the network. The cloud computing nodes then perform complex computational processing, such as neural network training and big data analysis. If the data already exists in the historical database, it remains in the temporary storage point. The edge nodes can perform simple data analysis and processing, such as data statistics and anomaly detection, to reduce the amount of data transmitted to the cloud and alleviate the computing pressure on the cloud platform.

[0081] Through the above specific implementation steps, efficient communication and intelligent processing of massive satellite data in the "air-space-ground-sea" integrated communication network have been achieved, verifying the feasibility and effectiveness of the method of the present invention.

[0082] To keep the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0083] Although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for 6G cloud-edge-end edge intelligent satellite internet combined with low earth orbit communication satellite data intelligence, characterized in that, Comprise the following steps: S1, in the satellite internet, constellation of "space-ground-sea" fusion communication network deployment laser emitting device, and through the AI analysis and prediction of power trend of control instruction; S2, build cloud-edge-end edge intelligent framework, the massive satellite data of receiving end adopts the optimal model segmentation method to form the grid of control command and contains multiple subgrids, and the layered model aggregation strategy is adopted between the grids.

2. The method for 6G cloud-edge-end edge intelligent satellite internet combined with low earth orbit communication satellite data intelligence according to claim 1, characterized in that, Step S1 specifically comprises: S1-1, for the different laser communication needs between satellite internet, constellation and opposite terminal, three kinds of laser inter-satellite links are established to support laser emitting device, and laser communication ranging integrated structure is adopted to process and control laser data; S1-2, laser emitting device is deployed in each part of the space-based satellite communication network, air-based satellite communication network, ground communication network and deep sea communication network of "space-ground-sea" fusion communication network; S1-3, through AI analysis and prediction of power trend of control instruction, and then quantize the required transmitting power of each network.

3. The method of claim 2, wherein the method is characterized by, The three kinds of laser inter-satellite links in step S1-1 are specifically: Laser communication between ground terminal and geosynchronous orbit platform GEO-GEO; Laser communication between ground terminal and geosynchronous orbit platform, low earth orbit platform GEO-LEO; Laser communication between ground terminal and low earth orbit platform LEO-LEO.

4. The method of claim 2, wherein the method is characterized by, The laser communication ranging integrated structure in step S1-1 is composed of transmitting subsystem, receiving subsystem, optical antenna, ATP subsystem and data processing and control subsystem; the system loads the information to be transmitted to the exciter of the modulator through the modulator, the exciter current of the modulator changes with the change law of the signal, and the output signal of the laser is modulated by the modulator, and the relevant parameters change according to the corresponding law; the optical antenna transmits the output signal of the laser, the detector detects the laser signal, and the original information is recovered through the demodulator to complete the transmission of data.

5. The method of claim 2, wherein the method is characterized by, In step S1-3, the power trend of control instruction is predicted by AI analysis, specifically, Markov transition matrix method is adopted to predict the power energy trend of next control instruction of the same type according to the required power of this time transmitting end control instruction; the model formula of Markov transition matrix method is: X(k+1)=X(k)×P Wherein, X(k) represents the state vector of the trend analysis and prediction object at t=k time, P represents one-step transition probability matrix, and X(k+1) represents the state vector of the trend analysis and prediction object at t=k+1 time.

6. The method of claim 5, wherein the method is characterized by, In step S1-3, large model is also used for fine tuning and Markov prediction for differential calibration, and the large model fine tuning specifically comprises: One-stage fine tuning, analyzing the data of control instruction arriving at ground terminal, calculating the loss loss of ridge regression overfitting, quantifying the power required by the loss part, and preliminarily referring to the power adjustment of transmitting end, the formula for preventing overfitting operation of ridge regression overfitting loss is: where L(θ) is the loss function; y i is the actual value of the control instruction power; X i is the input parameter matrix, including the communication capacity and the communication distance; θ is the fitting hyperparameter; α is the weight of the unit matrix, and the value range is 0.01-0.1, which is adjusted according to the data fluctuation degree; ||θ|| 2 is the L2 regularization term, which is used to prevent overfitting; According to the calculated L(θ), when L(θ) is greater than the preset threshold, the required transmitting power is calculated as power compensation value=L(θ)×reference power coefficient, which is used as the preliminary basis for transmitting end adjustment; Two-stage fine-tuning, the filtered log data set composed of the question set is used as the computing task, and the perplexity PPL is calculated after the computing power is quantified: where N is the length of the text sequence, w i is the i-th word in the sequence, p(w i |w 1:i-1 ) is the probability of the i-th word given the preceding words.

7. The method for 6G cloud-edge-end edge intelligent satellite internet combined with low-orbit communication satellite data intelligence according to claim 6, characterized in that, According to the demand, the large model fine-tuning can be expanded to three-stage, four-stage and multi-stage fine-tuning; If the multi-stage PPL fluctuation is less than 10%, the final PPL is set by weighted average; if the fluctuation is greater than 10%, the adjustment result of the stage before the fluctuation is retained.

8. The method of claim 1, wherein the method is used for 6G cloud-edge-end edge intelligent satellite internet combined with low earth orbit communication satellite data intelligence algorithm. The optimal model segmentation method in step S2 specifically includes: The data segmented by edge intelligent grid is used to analyze the data usage density of each control sub-grid, and the network usage density is calculated using the dynamically changing control command quantity, data flow and connection number, and the calculation formula is: Composite Density = w u • User Density + w t • Traffic Density + w c • Connection Density wherein Composite Density is the composite density of the command subnetwork, User Density is the density of the number of command grids, Traffic Density is the density of the subgrids, Connection Density is the density of the number of subgrid connections, w u , w t , and w c are weights for the number of subgrids, command, and number of connections, respectively; According to the sub-network usage density, the grid size of each sub-grid is adjusted, the high-density area is divided into larger grids, and the low-density area is divided into smaller grids, and the calculation formula is: Wherein, New Grid Size is the adjusted grid size, Base Grid Size is the initial grid size, Network Density is the network usage density of each sub-grid, and Average Density is the average network usage density of the entire grid.

9. The method according to claim 1, wherein, After the cloud-edge-end edge intelligent framework is constructed in step S2, when the same type of control instruction related data in the massive satellite data is stored in the cloud, the grid composed of the temporary storage network of the satellite edge node is used to relieve the computing pressure of the cloud platform; the same type of control instruction related data forms a large grid, which includes multiple sub-grids; if it is new data, it is transferred from the temporary storage point to the cloud storage and operation, otherwise it is left in the temporary storage point.

10. The method of claim 1, wherein the method is used for 6G cloud-edge-end edge intelligent satellite internet combined with low earth orbit communication satellite data intelligence algorithm. The parameters of each NTN node in the "space-air-ground-sea" integrated communication network are as follows: GEO satellite: altitude 35786km, power-Markov chain model is used to predict power trend of 30% increase and 70% decrease, power-large model fine-tuning test is used to predict power trend of 30% increase and 70% decrease with accuracy of 90%, and expected use cases include wide coverage, media broadcast, public content broadcast, fixed / mobile cell backhaul connection, urban / remote user communication; LEO satellite: altitude 400-1600km, expected use cases include fixed / mobile cell backhaul connection, urban / remote user communication, media multicast service, wide-area Internet of Things service; VLEO satellite: altitude 100-400km, expected use cases include mobile cell backhaul connection, urban / remote user communication, media multicast service, wide-area Internet of Things, broadband Internet; HAPS: altitude 15-25km, expected use cases include air / ground base station backhaul connection, urban / remote user communication, media multicast service, landing Internet of Things service; UAV: altitude 0.1-10km, expected use cases include ground base station backhaul connection, air relay / TRP, on-demand hotspot, regional emergency service.

Images