Star-ground data source packet unpacking and framing method, system, device and medium

Abstract

The application relates to the field of satellite communication and discloses a satellite-to-ground uplink data source packet unpacking and framing method, system, equipment and medium, which comprises the following steps: preprocessing: collecting channel environment data, satellite equipment state data and data source packet service attribute data, and preprocessing, wherein the data source packet service attribute data comprises correlation identification among data; feature extraction: using a multilayer nested network to extract features of the preprocessed data to obtain a comprehensive feature vector, including features of the data source packet and features of the correlation relationship among data; decision unpacking: inputting the comprehensive feature vector into a decision model fused by deep reinforcement learning and model predictive control to generate an unpacking decision. The satellite-to-ground uplink data source packet is collected and preprocessed, the multilayer nested network is used to extract features of the preprocessed data, the comprehensive feature vector is input into the decision model fused by deep reinforcement learning and model predictive control to generate the unpacking decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite communication, specifically a method, system, device and medium for satellite-ground uplink data source packet unpacking and framing. BACKGROUND

[0002] In the field of satellite communication today, with the increasing widespread application of satellites, such as global positioning, remote sensing monitoring, satellite communication network construction, the amount of data transmission between satellites and the ground is growing explosively. Efficient and stable transmission of various data source packets between satellites and ground stations is required, and these data packets contain different types of information, such as remote sensing image data, satellite device status data, control instruction data, etc. At the same time, satellite communication faces complex channel environments, including signal attenuation, multipath effect, noise interference, etc., as well as the challenges of limited satellite resources, such as limited storage capacity and computing power. This requires the continuous upgrading of processing technology for satellite-ground uplink data source packets to meet the growing communication needs and complex application scenarios.

[0003] Currently, in the field of satellite-ground uplink data source packet unpacking and framing, traditional methods usually use fixed unpacking and framing strategies. In the unpacking stage, data packets are often split according to pre-set fixed rules, such as fixed byte numbers or fixed formats, without considering real-time changes in channel conditions and differences in data source packet characteristics. In the framing process, data packets are simply combined in a certain order, lacking deep consideration of data correlation and transmission efficiency. The fixed unpacking and framing strategy lacks dynamic adaptability to channel conditions and data source packet characteristics, resulting in low data transmission efficiency. SUMMARY

[0004] To address the above problems, the present application provides a method, system, device and medium for satellite-ground uplink data source packet unpacking and framing, aiming to improve and solve the problem of low data transmission efficiency caused by the lack of dynamic adaptability of traditional satellite-ground uplink data source packet unpacking and framing methods to channel conditions and data source packet characteristics.

[0005] In a first aspect, the present application provides a method for satellite-ground uplink data source packet unpacking and framing, comprising the following steps:

[0006] Preprocessing: Collect channel environment data, satellite device status data and data source packet service attribute data, and preprocess, wherein the data source packet service attribute data contains association identifiers between data, and the association identifiers include associated data packet IDs and association type codes;

[0007] Feature extraction: using a multi-layer nested network to extract features from the preprocessed data to obtain a comprehensive feature vector, including the features of the data source packet and the features of the association relationship between data;

[0008] Decision unpacking: inputting the comprehensive feature vector into a decision model combined with deep reinforcement learning and model predictive control to generate unpacking decisions;

[0009] Frame generation: constructing a hypergraph neural network according to the association identifier and the feature representation of the association relationship between data, and generating a framing strategy combined with a multi-objective optimization algorithm;

[0010] Platform execution: executing unpacking and framing operations according to unpacking decisions and framing strategies under the scheduling of a real-time operating system through a heterogeneous hardware platform composed of FPGA, ASIC and GPU.

[0011] By adopting the above technical solutions, accurate information is obtained by collecting and preprocessing channel environment data, on-board device state data and data source packet service attribute data, comprehensive feature vectors are extracted through a multi-layer nested network to mine data features and association relationships, and a combined decision model and hypergraph neural network are used in combination with a multi-objective optimization algorithm to generate unpacking decisions and framing strategies that meet actual needs. With the help of a heterogeneous hardware platform and a real-time operating system, unpacking and framing operations are quickly and accurately completed, and the unpacking and framing strategies are dynamically adjusted according to real-time changes in channel state, the characteristics of data source packets and the status of on-board device resources. When the channel quality is good, the unpacking granularity is increased, the packet header overhead is reduced, and the transmission efficiency is improved. When the channel quality is poor, the unpacking granularity is reduced, the retransmission probability is reduced, and at the same time, the data source packets are reasonably framed according to the association relationship, improving the completeness and accuracy of data transmission. Thus, the problem of low data transmission efficiency caused by the lack of dynamic adaptability of the fixed unpacking and framing strategies of traditional satellite-to-ground data source packet unpacking and framing methods is solved.

[0012] Preferably, the preprocessing includes data cleaning, format conversion, standardization processing and timestamp synchronization, and the association identifier includes an associated data packet ID and an association type code.

[0013] Preferably, the feature extraction specifically includes the following steps:

[0014] The spatial features of the channel environment data are extracted using a convolutional neural network, and the formula is: wherein is the weight tensor of the lth convolution kernel, is the bias vector;

[0015] The time series features of the channel environment data are extracted using a stacked long short-term memory network.

[0016] The spatial features and the time sequence features are fused by the multi-head attention mechanism, and global feature integration is performed by combining knowledge graph embedding to obtain a comprehensive feature vector.

[0017] Preferably, the state space of the deep reinforcement learning includes a multi-modal deep feature vector, a dynamic change sequence of a current channel transmission rate, a multi-dimensional distribution of a residual storage resource of a satellite, a multi-dimensional state trend of a satellite device, and a service demand prediction of a data source packet, and the model predictive control includes establishing a channel transmission model, a satellite resource consumption model, and a service demand evolution model to predict state changes at future time steps.

[0018] Preferably, the reward function of the decision model is R(s, a) = ω1P success -ω2D+ω3U+ω4E+ω5S satisfaction ; wherein ω1+ω2+ω3+ω4+ω5=1; P success is a transmission success rate, D is a delay, U is a bandwidth utilization rate, E is a resource balance degree, and S is a service satisfaction degree, the generated unpacking decision is an initial action sequence generated by deep reinforcement learning, and the initial action sequence is optimized by model predictive control to minimize a cumulative loss function at future time steps, thereby generating.

[0019] Preferably, the frame group generation specifically includes the following steps:

[0020] Taking the data source packet as a node, determining the hyperedge connection relationship between nodes according to the association identifier in the data source packet service attribute data, including a service association hyperedge, a time sequence association hyperedge, and a resource association hyperedge;

[0021] calculating the association strength between nodes by the formula , wherein hi and hj are feature vectors of nodes i and j in the comprehensive feature vector, I(hi, hj) is mutual information, H(hi, hj) is joint information entropy, ω1+ω2+ω3=1, and the association strength is used as a hyperedge weight;

[0022] A three-layer HGNN network is built to perform feature extraction and association analysis on the hypergraph, the node feature dimension is set to 256, weighted summation is used for hyperedge aggregation, and a multi-objective optimization algorithm is used to generate a frame grouping strategy.

[0023] Preferably, the multi-objective optimization algorithm is a dynamic programming and a genetic algorithm, and the state transition equation of the dynamic programming is wherein A(i, z) is a feasible frame grouping action set for the first i data source packets under state z, d(a) is a resource vector consumed by action a, and g(j+1, i, a) is a cost vector of grouping the j+1th to ith data source packets according to action a.

[0024] Secondly, the present invention provides the following technical solution: a satellite-to-ground injection data source packet unpacking and framing system, used in the aforementioned satellite-to-ground injection data source packet unpacking and framing method, the system comprising:

[0025] Preprocessing module: used to collect channel environment data, on-board equipment status data and data source packet service attribute data, and perform preprocessing. The data source packet service attribute data includes the association identifier between data.

[0026] Feature extraction module: Used to extract features from preprocessed data using a multi-layer nested network to obtain a comprehensive feature vector, including features of the data source package and features of the relationships between data.

[0027] Decision unpacking module: used to input the comprehensive feature vector into a decision model that integrates deep reinforcement learning and model prediction control to generate unpacking decisions;

[0028] Frame generation module: Used to construct a hypergraph neural network based on the characteristics of association identifiers and relationships between data, and combine it with a multi-objective optimization algorithm to generate a frame generation strategy;

[0029] Platform execution module: Used to perform unpacking and framing operations through a heterogeneous hardware platform composed of FPGA, ASIC and GPU, under the scheduling of a real-time operating system, based on unpacking decisions and framing strategies.

[0030] Thirdly, the invention provides the following technical solution: a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned method for unpacking and framing satellite-to-ground injection data source packets.

[0031] Fourthly, the present invention provides the following technical solution: a readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned method for unpacking and framing satellite-to-ground injection data source packets.

[0032] This invention provides a method, system, device, and medium for unpacking and framing satellite-to-ground injection data source packets. It offers the following advantages:

[0033] 1. This invention collects and preprocesses satellite-to-ground data packets, extracts features from the preprocessed data using a multi-layer nested network, inputs the comprehensive feature vectors into a decision model that integrates deep reinforcement learning and model predictive control to generate packet splitting decisions, constructs a hypergraph neural network based on data correlations and combines it with a multi-objective optimization algorithm to generate a framing strategy, and executes packet splitting and framing operations. This allows for dynamic perception of changes in the satellite-to-ground communication environment and the characteristics of data packets. Based on the real-time channel status, satellite resource status, and the service requirements and correlations of data packets, the packet splitting and framing strategies are dynamically adjusted, solving the problem of low data transmission efficiency in traditional satellite-to-ground data packet splitting and framing methods.

[0034] 2. By collecting and analyzing on-board equipment status data, this invention can rationally arrange unpacking and framing tasks based on the satellite's storage and computing resources. When satellite storage resources are scarce, the decision model will adjust the unpacking strategy to avoid generating excessively large data blocks that consume too much storage resources. When computing resources are limited, the framing strategy will prioritize simple tasks to ensure stable system operation. In addition, the extraction of association identifiers and related features from the data source packet business attribute data can identify and process different types of data associations. Even when facing complex business scenarios, it can still accurately perform unpacking and framing operations, enhancing its adaptability to complex environments and diverse businesses, and improving the reliability of data transmission.

[0035] 3. This invention utilizes a deep reinforcement learning-based decision model that considers the multidimensional distribution of satellite resources. By predicting future resource consumption through model-based predictive control, it optimizes the data processing flow during the packet unpacking decision-making process, avoiding excessive resource occupation or waste. In the frame generation stage, a multi-objective optimization algorithm takes resource balance as an important objective. Through dynamic programming and genetic algorithms, it ensures that various satellite resources are rationally allocated and used in a balanced manner while meeting data transmission requirements, thereby improving the overall utilization efficiency of satellite system resources and extending the satellite's lifespan and operational stability.

[0036] 4. This invention utilizes the association identifiers in the business attribute data of data source packages and combines them with the features of data relationship between comprehensive feature vectors to construct a hypergraph neural network. This network can accurately capture complex business relationships, temporal relationships, and resource relationships between data source packages. By calculating the relationship strength between nodes and using it as the hyperedge weight, and by constructing a three-layer HGNN network for feature extraction and relationship analysis, this invention can deeply understand the intrinsic connections between data. As a result, these relationships are fully considered during the frame generation process, making the frame generation results more reasonable, effectively ensuring the integrity and accuracy of the data, and meeting the strict requirements for data processing accuracy in complex business scenarios. Attached Figure Description

[0037] Figure 1 This is a flowchart of the satellite-to-ground injection data source packet unpacking and framing method proposed in this invention;

[0038] Figure 2 This is a system architecture diagram of the satellite-to-ground injection data source packet unpacking and framing system proposed in this invention. Detailed Implementation

[0039] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1

[0041] Reference Figure 1 In the first embodiment of the present invention, the present invention provides a method for unpacking and framing satellite ground injection data source packets, including the following steps:

[0042] Preprocessing: Collect channel environment data, on-board equipment status data, and data source packet service attribute data, and perform preprocessing. The data source packet service attribute data contains the association identifiers between data. Preprocessing includes data cleaning, format conversion, standardization, and timestamp synchronization. The association identifiers include the associated data packet ID and the association type code.

[0043] Specifically, data is collected through various sensors and monitoring equipment. The ground control center uses multi-band radar to collect channel environment data, covering signal strength, frequency offset, and other information across different frequency bands, to monitor channel stability and interference. A meteorological sensor network acquires meteorological data such as temperature, humidity, and air pressure, as weather conditions affect signal propagation quality. Satellite monitoring equipment tracks the satellite's orbital position, attitude, and other parameters in real time to ensure the satellite is in normal working order. At the satellite end, storage status sensors, computing resource sensors, and environmental parameter sensors installed on the satellite payload collect data on satellite storage resource usage, computing chip load, and the temperature, vibration, and pressure of the satellite's environment, respectively. When data source packets are generated, they carry business attribute data, including associated data packet IDs to clarify the relationship between this data packet and other data packets, as well as association type codes, such as business association, time-series association, and resource association. These codes are used to distinguish different association types.

[0044] Outliers are removed using the 3σ principle. For channel transmission rate data, if a measurement deviates from the mean by more than three times the standard deviation, it is considered an outlier and removed from the dataset. This is because outliers may be caused by sensor malfunctions, interference, or other factors that can affect the accuracy of subsequent data processing. When processing satellite storage resource usage data, data exhibiting sudden, large fluctuations that deviate from the normal range are also cleaned according to this principle.

[0045] All collected data in various formats are uniformly converted to JSON format. For example, binary data collected by sensors and business attribute data in XML format are all converted to JSON format using appropriate conversion programs. JSON format has good readability and universality, which facilitates subsequent data parsing and processing, and also makes it easier for different systems and modules to share and exchange data.

[0046] Channel data is normalized to map its numerical range to the [0,1] interval. For satellite equipment status data, the Min-Max normalization method is used to adjust the data to a specific range, eliminating the influence of different units of measurement. Taking satellite computing resource utilization and storage resource utilization as examples, Min-Max normalization unifies them to the same numerical scale, facilitating subsequent data analysis and model training.

[0047] By utilizing high-precision time synchronization systems such as GPS or BeiDou, the timestamps of various data types are uniformly calibrated to ensure time accuracy within ±100μs. This ensures consistency and accuracy of data from different sources in the time dimension, providing a reliable foundation for subsequent time-series data analysis and processing. In channel state analysis, accurate time synchronization helps determine the sequence and correlation of signal changes.

[0048] Feature extraction: A multi-layered nested network is used to extract features from the preprocessed data, obtaining a comprehensive feature vector, including features of the data source package and features of the relationships between data points. The feature extraction process specifically includes the following steps:

[0049] Spatial features are extracted from channel environment data using a convolutional neural network, and the formula is as follows: in Let be the kernel weight tensor of the ι-th layer. It is the bias vector;

[0050] A stacked long short-term memory network is used to extract time-series features from channel environment data to obtain time-series features;

[0051] By fusing spatial and temporal features through a multi-head attention mechanism and integrating global features through knowledge graph embedding, a comprehensive feature vector is obtained.

[0052] Specifically, convolutional neural networks (CNNs) are used to extract spatial features from channel environment data. Taking a four-layer convolutional architecture as an example, the first convolutional layer uses 32 3×3 convolutional kernels with a stride of 1 to perform convolution operations on the input channel environment data. The convolutional kernels slide across the data, extracting local features by multiplying with and summing data elements. These features are then enhanced by the ReLU activation function to improve their non-linear expressiveness. The second convolutional layer uses 64 3×3 convolutional kernels to further extract higher-level spatial features. Next, a 2×2 max-pooling layer downsamples the data, reducing its dimensionality while preserving key features. Finally, a fully connected layer maps the pooled features to a 128-dimensional spatial feature vector. (Formula follows) The computation process of convolutional layers is described, demonstrating how input data is transformed using convolutional kernels and bias vectors to extract spatial features at different levels.

[0053] A Stacked Long Short-Term Memory (LSTM) network is employed to extract temporal features from channel environment data. A three-layer LSTM network is constructed, with each layer containing 256 neurons. LSTM units control the flow and retention of information through input gates, forget gates, and output gates. When processing time-series data, it can learn long-term dependencies. For example, when analyzing changes in channel transmission rate over time, the LSTM network can remember past rate trends and make comprehensive judgments based on current input data, thereby extracting more accurate temporal features. The computation process of each LSTM unit is as follows:

[0054] i t =σ(W xi E t +W hi h t-1 +b i )

[0055] f t =σ(W xf E t +W hf h t-1 +b f )

[0056] c t =f t ⊙c t-1 +i t ⊙tanh(W xc E t +W hc H t-1 +b c )

[0057] o t =σ(Wxo E t +W ho h t-1 +b o )

[0058] ht=ot⊙tanh(c t );

[0059] Among them, i t f t o t These are the input gate, forget gate, and output gate, respectively. t Let be the cell state, ht be the hidden state, σ be the sigmoid function, ⊙ be the element-wise multiplication, W be the weight matrix, and b be the bias vector. Through this calculation process, the LSTM network can effectively process time-series data and extract the time-dimensional features of channel environment data.

[0060] Spatial and temporal features are fused using a multi-head attention mechanism. Eight attention heads are used, each with a dimension of 64. This multi-head attention mechanism focuses on different aspects of the data through different attention heads, enabling a more comprehensive capture of the relationships between features. Attention weights are calculated using ScaledDot-ProductAttention, with the following formula: (in and Let d be the query and key weight matrix for the m-th head. k (This is the key vector dimension). In this way, the similarity between different feature vectors is calculated, and attention weights are assigned based on the similarity, fusing spatial and temporal features. Then, global feature integration is performed by combining knowledge graph embedding. A knowledge graph containing satellite communication-related entities (such as satellites, ground stations, communication protocols, etc.) and relationships (such as connection relationships, dependencies, etc.) is constructed. The TransE algorithm is used for knowledge representation learning, embedding information from the knowledge graph into the feature vectors to obtain a comprehensive feature vector. Thus, the comprehensive feature vector not only includes the features of the data itself but also incorporates prior knowledge in the field of satellite communication, more comprehensively reflecting the characteristics of the data source packets and the characteristics of the relationships between data.

[0061] Decision-making packet splitting: The comprehensive feature vector is input into a decision model that integrates deep reinforcement learning and model predictive control to generate packet splitting decisions. The state space of deep reinforcement learning includes multimodal deep feature vectors, dynamic change sequences of the current channel transmission rate, multidimensional distribution of satellite remaining storage resources, multidimensional state trends of onboard equipment, and service demand prediction of data source packets. Model predictive control includes establishing a channel transmission model, a satellite resource consumption model, and a service demand evolution model to predict state changes in several future time steps.

[0062] The reward function of the decision model is R(s,a)=ω1P success -ω2D+ω3U+ω4E+ω5S satisfaction ;where ω1+ω2+ω3+ω4+ω5=1;P success Let D be the transmission success rate, U be the latency, E be the bandwidth utilization, E be the resource balance, and S be the service satisfaction. The packet splitting decision is generated by generating an initial action sequence through deep reinforcement learning, and then using model predictive control to optimize the initial action sequence to minimize the cumulative loss function over several future time steps, thereby generating the packet.

[0063] Specifically, a decision-making model integrating deep reinforcement learning and model predictive control is employed. The deep reinforcement learning component constructs a multi-layer neural network. The input layer receives a comprehensive feature vector, which passes through two hidden layers, each containing 256 neurons. Finally, the output layer outputs the action (i.e., the packet splitting decision, such as the granularity and order of packet splitting). During training, a learning rate of 0.0001 and a discount factor of 0.99 are set. An experience replay mechanism is used to store and reuse training data. The experience replay buffer has a capacity of 100,000 records, and 64 records are randomly selected from the buffer for training each time to improve training efficiency and stability.

[0064] The state space of deep reinforcement learning includes multimodal deep feature vectors, a dynamic sequence of current channel transmission rate changes, a multidimensional distribution of remaining satellite storage resources, multidimensional state trends of onboard devices, and service demand predictions for data source packets. The multimodal deep feature vectors contain various feature information extracted from the data; the dynamic sequence of current channel transmission rate changes records recent channel transmission rate variations, used to determine channel stability; the multidimensional distribution of remaining satellite storage resources reflects the remaining capacity of different storage areas on the satellite, enabling rational data storage allocation; the multidimensional state trends of onboard devices demonstrate changes in the operating status of devices (such as computing chips, sensors, etc.); and the service demand predictions for data source packets estimate the processing requirements of data source packets based on historical data and business logic, such as predicting whether a particular data source packet needs to be processed first.

[0065] The model predictive control section establishes a channel transmission model, a satellite resource consumption model, and a service demand evolution model. The channel transmission model uses the ARIMA(3,1,2) model, which predicts the channel transmission state at several future time steps (e.g., 10 time steps) by analyzing and fitting historical channel data, including indicators such as signal strength and bit error rate. The satellite resource consumption model utilizes an LSTM network to predict future resource consumption trends based on the current satellite resource usage and data processing tasks. The service demand evolution model predicts future changes in service demand based on service priority, data volume, and historical processing data. These models predict state changes at multiple future time steps, providing a basis for decision-making.

[0066] The reward function of the decision model is R(s,a)=ω1P success -ω2D+ω3U+ω4E+ω5S satisfaction ;where ω1+ω2+ω3+ω4+ω5=1;P success For transmission success rate, for latency, for bandwidth utilization, for resource balance, for service satisfaction,

[0067] It is an important indicator for measuring the quality of decisions; the higher the transmission success rate, the greater the reward; D is latency, representing the time delay from data transmission to reception; the lower the latency, the higher the reward; U is bandwidth utilization, reflecting the efficiency of communication resource utilization; the higher the utilization rate, the higher the reward; E is resource balance, used to measure the degree of balanced use of satellite resources (such as storage resources, computing resources); the more balanced the resource use, the higher the reward; S satisfaction To assess business satisfaction, evaluations are based on how well business needs are met; the higher the degree of satisfaction, the higher the reward. By adjusting the weighting coefficients ω1, ω2, ω3, ω4, and ω5, the focus of the decision-making model can be adjusted according to different business scenarios and needs. For example, in scenarios with high real-time requirements, ω2 (the weight of latency) can be appropriately increased to prioritize reducing data transmission latency.

[0068] An initial action sequence is generated using deep reinforcement learning, containing preliminary packet splitting decisions. Then, model predictive control (MMCC) is used to optimize this initial action sequence. MMCC predicts future state changes, calculates the cumulative loss function for each action sequence over multiple time steps, and selects the optimal action sequence by minimizing the cumulative loss function, thus generating the final packet splitting decision. This decision-making process comprehensively considers both the current state and future trends, resulting in more reasonable packet splitting decisions.

[0069] Frame generation: Based on the association identifiers and the characteristics of the relationships between data, a hypergraph neural network is constructed, and a multi-objective optimization algorithm is used to generate a frame generation strategy; the frame generation specifically includes the following steps:

[0070] Using data source packages as nodes, determine the hyperedge connection relationships between nodes based on the association identifiers in the business attribute data of the data source packages, including business association hyperedges, time-series association hyperedges, and resource association hyperedges;

[0071] Through formula Calculate the association strength between nodes, where hi and hj are the feature vectors of nodes i and j in the comprehensive feature vector, I(hi,hj) is the mutual information, H(hi,hj) is the joint information entropy, ω1+ω2+ω3=1, and use the association strength as the hyperedge weight;

[0072] A three-layer HGNN network was constructed to extract features and perform correlation analysis on the hypergraph. The node feature dimension was set to 256. Weighted summation was used for hyperedge aggregation, and a multi-objective optimization algorithm was combined to generate a frame grouping strategy.

[0073] Multi-objective optimization algorithms include dynamic programming and genetic algorithms. The state transition equation for dynamic programming is: ; where A(i,z) is the set of feasible frame-building actions for the first i data source packets in state z, d(a) is the resource vector consumed by action a, and g(j+1,i,a) is the cost vector for framing the j+1 to i data source packets according to action a.

[0074] Specifically, using data source packets as nodes, the hyperedge connections between nodes are determined based on the association identifiers in the business attribute data of the data source packets. When two data source packets are related in terms of business, such as belonging to different parts of the same task, a business association hyperedge is established; if there is a temporal sequence between the data source packets, such as the processing result of one data packet being the input of another data packet, a temporal association hyperedge is established; if the data source packets are related in terms of resource usage, such as both relying on the same satellite storage area or computing resources, a resource association hyperedge is established.

[0075] Through formula Calculate the association strength between nodes (where hi and hj are the feature vectors of nodes i and j in the comprehensive feature vector, respectively, I(hi,hj) is the mutual information, H(hi,hj) is the joint information entropy, and ω1+ω2+ω3=1). Cosine similarity. The similarity between two feature vectors is measured; mutual information I(hi,hj) represents the degree of information sharing between two variables, reflecting their dependency; joint information entropy H(hi,hj) describes the uncertainty between the two variables. By combining these three indicators and weighting them according to different weight coefficients ω1, ω2, and ω3, a more accurate association strength value is obtained. Association strength is used as the hyperedge weight; the larger the weight, the stronger the association between the two data source packets.

[0076] A three-layer HGNN network was constructed, with the node feature dimension set to 256. In the hypergraph neural network, feature extraction and association analysis were performed on the hypergraph through hypergraph convolution operations. Weighted summation was used for hyperedge aggregation, and the node features connected by hyperedges were aggregated according to their weights to obtain more representative node feature representations.

[0077] A multi-objective optimization algorithm is used to generate a frame grouping strategy, employing a combination of dynamic programming and genetic algorithms. The state transition equation for dynamic programming is as follows: Where A(i,z) is the set of feasible framing actions for the first i data source packets in state z, d(a) is the resource vector consumed by action a, and g(j+1,i,a) is the cost vector for framing the (j+1)th to the ith data source packets according to action a. Dynamic programming finds the scheme that minimizes the framing cost while satisfying resource constraints by solving subproblems step by step. Genetic algorithm is used for global optimization. It simulates the biological evolution process and iteratively optimizes the framing scheme through operations such as selection, crossover, and mutation. In the genetic algorithm, the framing scheme is encoded as chromosomes, the population size is set to 100, the crossover probability is 0.8, the mutation probability is 0.01, and the selection strategy is tournament selection (size 5). By continuously evolving the population, a better framing strategy is found, comprehensively considering multiple objectives such as transmission delay, bandwidth utilization, data integrity, and service priority.

[0078] Platform execution: Through a heterogeneous hardware platform composed of FPGA, ASIC and GPU, under the scheduling of the real-time operating system, the unpacking and framing operations are performed according to the unpacking decision and framing strategy.

[0079] Specifically, the Xilinx Zynq UltraScale+MPSoC was chosen as the FPGA module, integrating an ARM processor core and abundant programmable logic resources. Its programmable logic section is used to implement flexible packet unpacking and framing logic, programmed using hardware description languages ​​(such as Verilog or VHDL) to achieve rapid processing of data source packets. For example, a finite state machine (FSM) is designed to control the unpacking and framing process, executing corresponding operations based on different states. The ARM processor core is used to manage and coordinate the FPGA's operation and communicate with other hardware modules.

[0080] An ASIC module designed using a 7nm FinFET process is customized to perform computationally intensive tasks such as CRC checksum, AES encryption, and data compression. The ASIC chip's CRC checksum unit supports 16 / 32-bit checksum, enabling fast and accurate error detection during data transmission; the AES encryption unit achieves a throughput of up to 10Gbps, providing strong encryption protection for data transmission; and the data compression unit employs the LZ77 algorithm, effectively reducing data transmission volume and improving transmission efficiency.

[0081] A GPU cluster was constructed, with each node equipped with eight NVIDIA A100 GPUs, interconnected via an InfiniBand HDR 200G network. Leveraging the powerful parallel computing capabilities of GPUs, the inference computation of deep learning models, such as deep reinforcement learning models and hypergraph neural networks, was accelerated. During hypergraph neural network computation, GPUs can simultaneously handle feature calculations from multiple nodes, significantly improving computational speed. Simultaneously, high-speed data transmission via the InfiniBand network ensured efficient collaboration between the GPU nodes.

[0082] Task scheduling is performed using the real-time operating system (RTOS) μC / OS-III. μC / OS-III supports 256 priority levels and employs a hybrid scheduling algorithm combining EDF (Earliest Deadline First) and RM (Rate Monotonic) algorithms. For time-critical tasks, such as high-priority data packet processing, the EDF algorithm is used to arrange the execution order based on the task's deadline, ensuring timely completion. For periodic tasks, such as data acquisition, the RM algorithm is used to allocate priorities based on the task's cycle, ensuring efficient use of system resources. μC / OS-III's interrupt response time is no more than 10μs, enabling rapid response to external events such as hardware failures or the arrival of new data, guaranteeing the system's real-time performance and stability.

[0083] Under the scheduling of the real-time operating system, operations are executed according to packet splitting decisions and framing strategies. When a data source packet is received, the FPGA determines the granularity and order of packet splitting based on the packet splitting decision, breaking the data source packet into data blocks of appropriate size. The ASIC performs CRC checks, encryption, and decryption on the split data blocks to ensure data integrity and security. The GPU leverages its powerful computing capabilities to accelerate the inference process of the deep learning model, providing support for packet splitting decisions and frame generation. The processed data blocks are then framed according to the framing strategy to form new data packets, which are transmitted via the satellite communication link.

[0084] Example 2:

[0085] Reference Figure 2In a second embodiment of the present invention, the present invention provides a satellite-to-ground injection data source packet unpacking and framing system for the above-described satellite-to-ground injection data source packet unpacking and framing method. The system includes:

[0086] Preprocessing module: Used to collect channel environment data, on-board equipment status data and data source packet service attribute data, and perform preprocessing. The data source packet service attribute data contains the correlation identifier between data.

[0087] Feature extraction module: Used to extract features from preprocessed data using a multi-layer nested network to obtain a comprehensive feature vector, including features of the data source package and features of the relationships between data.

[0088] Decision unpacking module: This module takes the comprehensive feature vector as input to a decision model that integrates deep reinforcement learning and model predictive control, and generates unpacking decisions.

[0089] Frame generation module: Used to construct a hypergraph neural network based on the characteristics of association identifiers and relationships between data, and combine it with a multi-objective optimization algorithm to generate a frame generation strategy;

[0090] Platform execution module: Used to perform unpacking and framing operations through a heterogeneous hardware platform composed of FPGA, ASIC and GPU, under the scheduling of a real-time operating system, based on unpacking decisions and framing strategies.

[0091] Specifically, the preprocessing module is responsible for collecting channel environment data, on-board equipment status data, and data source packet service attribute data, and performing preprocessing. This module integrates multiple sensor interfaces and data acquisition circuits, enabling real-time acquisition of various types of data. It employs a 3σ principle algorithm to detect and remove outliers from the collected data; the format conversion unit uses a dedicated conversion program to convert data of different formats to JSON format; the standardization processing unit standardizes the data according to the characteristics of different data types using algorithms such as normalization and Min-Max; and the timestamp synchronization unit uses a high-precision time synchronization system (such as GPS or BeiDou) to accurately calibrate the data timestamps. Through these operations, the preprocessing module provides a high-quality data foundation for subsequent data processing.

[0092] Feature Extraction Module: This module utilizes a multi-layered nested network to extract features from the preprocessed data. The Convolutional Neural Network (CNN) section employs specially designed convolutional kernels and network structures to extract spatial features from the channel environment data. Through operations such as convolution, activation, and pooling, it progressively extracts local and global spatial features of the data. The Stacked Long Short-Term Memory (LSTM) network, through the stacking of multiple LSTM units, performs in-depth learning of the temporal features of the channel environment data, effectively capturing the dependencies of data over time. The multi-head attention mechanism fusion unit calculates attention weights to fuse the spatial features extracted by the CNN and the temporal features extracted by the LSTM. Combined with the knowledge graph embedding unit, it integrates knowledge from the satellite communication field into the feature vector, ultimately obtaining a comprehensive feature vector that includes data packet features and features related to the relationships between data.

[0093] The packet splitting decision module: This module inputs the comprehensive feature vector into a decision model that integrates deep reinforcement learning and model predictive control to generate packet splitting decisions. The deep reinforcement learning network employs a multi-layer neural network structure, continuously learning the optimal packet splitting strategy through training. The model predictive control part utilizes established channel transmission models, satellite resource consumption models, and service demand evolution models to predict future state changes and optimize the initial action sequence generated by deep reinforcement learning. The decision model evaluates different packet splitting decisions based on a reward function. This reward function comprehensively considers multiple indicators such as transmission success rate, latency, bandwidth utilization, resource balance, and service satisfaction. By adjusting the weights of the reward function, the packet splitting decision can be optimized according to different service scenarios and requirements.

[0094] The frame generation module constructs a hypergraph neural network based on association identifiers and the characteristics of relationships between data points, and generates a frame generation strategy using a multi-objective optimization algorithm. The hypergraph construction unit determines the hyperedge connections between nodes based on association identifiers in the business attribute data of the data source packets, including business-related hyperedges, temporal-related hyperedges, and resource-related hyperedges. The association strength calculation unit calculates the association strength between nodes using a specific formula and uses it as the hyperedge weight. The hypergraph neural network (HGNN) processing unit builds a three-layer HGNN network to perform feature extraction and association analysis on the hypergraph. It aggregates hyperedges through hypergraph convolution operations and weighted summation to learn the complex relationships between data source packets. The multi-objective optimization algorithm unit uses a combination of dynamic programming and genetic algorithms to generate the optimal frame generation strategy, aiming to minimize transmission latency, maximize bandwidth utilization, ensure data integrity, and meet business priorities.

[0095] Platform Execution Module: Utilizing a heterogeneous hardware platform comprised of FPGA, ASIC, and GPU, and under the scheduling of the real-time operating system, the module executes packet unpacking and framing operations based on packet unpacking decisions and framing strategies. The FPGA is responsible for implementing flexible packet unpacking and framing logic, achieving efficient data processing through hardware programming. The ASIC is customized to accelerate computationally intensive tasks such as CRC checksums, encryption / decryption, and data compression, improving processing speed and efficiency. The GPU leverages its powerful parallel computing capabilities to accelerate inference computations for deep learning models, providing support for decision-making and framing. The real-time operating system (such as μC / OS-III) is responsible for scheduling the tasks of each hardware module, ensuring the system's real-time performance and stability, and enabling fast and accurate processing of data source packets.

[0096] Example 3

[0097] In the third embodiment of the present invention, based on the same inventive concept, the present invention proposes a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the satellite-to-ground injection data source packet unpacking and framing method of the above embodiments.

[0098] Example 4

[0099] In the fourth embodiment of the present invention, based on the same inventive concept, a computer device is proposed, comprising: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory and execute the satellite-to-ground injection data packet unpacking and framing method of the above embodiment.

[0100] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0101] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for unpacking and framing data source packets from satellite ground injection, characterized in that, Includes the following steps: Preprocessing: Collect channel environment data, on-board equipment status data, and data source packet service attribute data, and perform preprocessing. The data source packet service attribute data includes the association identifier between data. Feature extraction: Multi-layer nested networks are used to extract features from the preprocessed data to obtain a comprehensive feature vector, including features of the data source package and features of the relationships between data. Decision unpacking: The comprehensive feature vector is input into a decision model that integrates deep reinforcement learning and model prediction control to generate unpacking decisions; Frame generation: Based on the characteristics of association identifiers and relationships between data, a hypergraph neural network is constructed, and a multi-objective optimization algorithm is used to generate a frame generation strategy; Platform execution: Through a heterogeneous hardware platform composed of FPGA, ASIC and GPU, under the scheduling of the real-time operating system, it performs unpacking and framing operations according to the unpacking decision and framing strategy; The feature extraction specifically includes the following steps: Spatial features are obtained by using convolutional neural networks to extract spatial features from channel environment data; A stacked long short-term memory network is used to extract time-series features from channel environment data to obtain time-series features; By fusing spatial and temporal features through a multi-head attention mechanism and integrating global features through knowledge graph embedding, a comprehensive feature vector is obtained. The generation and unpacking decision is to generate an initial action sequence through deep reinforcement learning, and then optimize the initial action sequence using model predictive control to minimize the cumulative loss function over several future time steps. The proposed frame generation strategy combines a multi-objective optimization algorithm with a combination of dynamic programming and genetic algorithms.

2. The method for unpacking and framing satellite ground injection data source packets according to claim 1, characterized in that: The preprocessing includes data cleaning, format conversion, standardization, and timestamp synchronization. The association identifier includes the associated data packet ID and the association type encoding.

3. The method for unpacking and framing satellite ground injection data source packets according to claim 1, characterized in that: The formula for the convolutional neural network is: .

4. The method for unpacking and framing satellite ground injection data source packets according to claim 1, characterized in that: The state space of the deep reinforcement learning includes multimodal deep feature vectors, dynamic change sequence of the current channel transmission rate, multidimensional distribution of satellite remaining storage resources, multidimensional state trends of on-board equipment, and service demand prediction of data source packets. The model prediction control includes establishing a channel transmission model, a satellite resource consumption model, and a service demand evolution model to predict state changes in several future time steps.

5. The method for unpacking and framing satellite ground injection data source packets according to claim 1, characterized in that: The reward function of the decision model is ;in ; To improve transmission success rate, For delay, For bandwidth utilization, For resource balance, For business satisfaction.

6. The method for unpacking and framing satellite ground injection data source packets according to claim 1, characterized in that: The frame generation specifically includes the following steps: Using data source packages as nodes, determine the hyperedge connection relationships between nodes based on the association identifiers in the business attribute data of the data source packages, including business association hyperedges, time-series association hyperedges, and resource association hyperedges; Through formula ; Calculate the association strength between nodes, where and They are nodes and The eigenvectors in the comprehensive eigenvectors For mutual information, For joint information entropy, And the correlation strength is used as the weight of the hyperedge; A three-layer HGNN network was constructed to perform feature extraction and association analysis on the hypergraph. The node feature dimension was set to 256. Weighted summation was used for hyperedge aggregation, and a multi-objective optimization algorithm was combined to generate a frame grouping strategy.

7. The method for unpacking and framing satellite ground injection data source packets according to claim 1, characterized in that: The state transition equation of the dynamic programming is: ;in In the state Down to front A set of feasible frame assembly actions for each data source packet. For action The vector of resources consumed. To make the first To the Each data source packet according to action The cost vector for framing.

8. A satellite-to-ground data source packet unpacking and framing system, characterized in that, The system for the satellite ground injection data source packet unpacking and framing method according to any one of claims 1-7 includes: Preprocessing module: used to collect channel environment data, on-board equipment status data and data source packet service attribute data, and perform preprocessing. The data source packet service attribute data includes the association identifier between data. Feature extraction module: Used to extract features from preprocessed data using a multi-layer nested network to obtain a comprehensive feature vector, including features of the data source package and features of the relationships between data. Decision unpacking module: used to input the comprehensive feature vector into a decision model that integrates deep reinforcement learning and model prediction control to generate unpacking decisions; Frame generation module: Used to construct a hypergraph neural network based on the characteristics of association identifiers and relationships between data, and combine it with a multi-objective optimization algorithm to generate a frame generation strategy; Platform execution module: Used to perform unpacking and framing operations under the scheduling of a real-time operating system through a heterogeneous hardware platform composed of FPGA, ASIC and GPU, based on unpacking decisions and framing strategies; The feature extraction specifically includes the following steps: Spatial features are obtained by using convolutional neural networks to extract spatial features from channel environment data; A stacked long short-term memory network is used to extract time-series features from channel environment data to obtain time-series features; By fusing spatial and temporal features through a multi-head attention mechanism and integrating global features through knowledge graph embedding, a comprehensive feature vector is obtained. The generation and unpacking decision is to generate an initial action sequence through deep reinforcement learning, and then optimize the initial action sequence using model predictive control to minimize the cumulative loss function over several future time steps. The proposed frame generation strategy combines a multi-objective optimization algorithm with a combination of dynamic programming and genetic algorithms.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the satellite ground injection data source packet unpacking and framing method as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which, when executed by a processor, implements the satellite ground injection data source packet unpacking and framing method as described in any one of claims 1 to 7.

Images