Satellite communication and edge network intelligent sensing method based on federated learning

Abstract

The invention discloses a federated learning-based satellite communication and edge network intelligent sensing method, which comprises the following steps of: 1, acquiring network edge environment, target and communication data, and preprocessing the data; 2, edge federated learning model training: performing fine tuning on pre-trained global model parameters at edge nodes by using local data, and transmitting the parameters to a base station through satellite or wireless communication for aggregation to obtain a new version; and 3, a non-orthogonal transmission and air computing technology design scheme is used for model parameter transmission of air computing. And 4, unloading a part of tasks to a base station, an edge server or a satellite according to the perception task complexity and time delay, realizing multi-dimensional calculation in the same resource block, improving the resource utilization rate and supporting satellite-ground cooperation. And 5, designing a joint beam forming algorithm to coordinate communication, perceive and calculate interference. And 6, based on network and service characteristics, the edge control center intelligently arranges services and coordinates ground and satellite network resources. And comprehensive acquisition of environment information, target states and communication data is realized.

Description

Technical Field

[0001] The present invention relates to the field of communications, and in particular to a satellite communication and edge network intelligent perception method based on federated learning. Background Art

[0002] In recent years, with the rapid development of artificial intelligence, Internet of Things technology and edge computing, machine learning algorithms have gradually played an important role in smart devices and services. Federated learning, as a new branch of machine learning, has brought revolutionary changes to this field.

[0003] Federated learning is a decentralized machine learning algorithm. Its basic principle is to integrate data from multiple edge devices to form a federated learning model to optimize the global model. In satellite communications and edge network environments, the number of terminal devices is huge and widely distributed, and the amount of data generated is huge and has strong privacy. Traditional machine learning methods require centralized storage and processing of data, which not only poses a risk of data privacy leakage, but also makes it difficult to meet the needs of decentralized data deployment. Federated learning can achieve efficient model training while protecting data privacy.

[0004] In satellite communications, due to the long communication links and high latency, traditional centralized machine learning algorithms are difficult to ensure real-time and accuracy. Federated learning greatly reduces communication overhead and latency by bringing the model directly to the terminal device for training and sending the trained parameters to the edge server for aggregation. Summary of the invention

[0005] To solve the above problems, the present invention proposes a satellite communication and edge network intelligent perception method based on federated learning. The specific steps are as follows:

[0006] Step 1: Multi-source data collection and preprocessing: deploy wireless sensors and cameras to cover the edge of the network, collect environmental information, target status and communication data, use low-orbit satellites or geostationary satellites to collect perception data in areas that cannot be covered by the ground, and preprocess the data;

[0007] Step 2: Edge federated learning model training: fine-tune the pre-trained global model using the local data set at the edge node, update the local model parameters, and upload the updated local model parameters to the base station via satellite communication or ground wireless communication. The base station uses edge federated learning technology to aggregate the model parameters and obtain a new version of the global model.

[0008] Step 3: Model parameter transmission based on air computing, using non-orthogonal transmission technology and air computing technology, design joint device selection and transceiver schemes to maximize the number of devices participating in federated learning and minimize model parameter transmission errors;

[0009] Step 4: Coordinate the perception and computing tasks. According to the complexity and latency requirements of the perception tasks, some perception tasks are offloaded to base stations, edge servers or satellites for computing, reducing the computing burden of edge nodes and ground networks, achieving multi-dimensional computing on the same resource block, improving resource utilization, and supporting the coordination of satellite communications and ground communications.

[0010] Step 5: Functional reuse and interference coordination: design a joint transmit and receive beamforming algorithm based on maximizing weighted total performance and minimizing total transmit power, coordinate interference between communication, perception, and computing, improve overall system performance, and support coordination between satellite and terrestrial communications.

[0011] Step 6: Intelligent business orchestration and optimization. Based on the global network perception and business characteristics, the edge control center performs intelligent business orchestration. According to the needs of different application scenarios, it dynamically adjusts the allocation and scheduling of communication, perception and computing resources, and coordinates the collaborative work of the ground network and the satellite network.

[0012] As a further improvement of the present invention, the multi-source data collection and preprocessing in step 1 can be expressed as:

[0013] Step 1.1: Ground sensing data collection: wireless sensors and cameras are deployed at the edge of the network. The sensors fully cover the target area through the layout of topological structure planning to obtain the most complete environmental information, target status and communication data;

[0014] Step 1.2, satellite sensing data collection, uses low-orbit satellites and geostationary satellites to obtain sensing data of areas that cannot be covered by the ground. Low-orbit satellites achieve continuous coverage through a network of multiple satellites. The relationship between the orbit period and orbit altitude of a single low-orbit satellite can be expressed as:

[0015]

[0016] Among them, T LEO is the orbital period of the satellite, h LEO is the track height, R E is the radius of the Earth, G is the gravitational constant, M E is the mass of the Earth;

[0017] Step 1.3, using Huffman coding compression for character data, so that characters with high frequency of occurrence are represented by shorter codes, and characters with low frequency of occurrence are represented by longer codes, thereby achieving character data compression;

[0018] Step 1.4, using the LZ77 algorithm for text data, by setting the sliding window and the look-ahead window, looking for the longest string in the look-ahead window that matches the string in the sliding window, if a matching string is found, using a pointer and length to represent the matching string;

[0019] Step 1.5: Use JPEG compression on the image data. First, divide the image into 8×8 blocks. Let the image be I(x, y). Each block can be represented by I i,j (x, y), i and j represent the position of the small block in the image, x and y represent the image coordinates, and a discrete cosine transform is performed on each small block:

[0020]

[0021] Among them, F(u,v) is the coefficient after discrete cosine transform, u and v are the frequency coordinates after transform; then the discrete cosine transform coefficient is quantized, and the quantized coefficient F q (u,v) can be expressed as:

[0022]

[0023] Among them, Q(u,v) is the quantization matrix, and finally the quantized coefficients are entropy encoded and encoded into compressed data.

[0024] As a further improvement of the present invention, the process of edge federated learning model training in step 2 can be expressed as follows:

[0025] Step 2.1: Local model training. Let the pre-trained global model be f(x;θ), where x is the input data and θ is the model parameter, and establish a local dataset. Where n is the number of local data samples, x i The i-th input sample, y i is the corresponding label;

[0026] Step 2.2, update the model parameters and use the cross entropy loss function to train the model. The cross entropy loss function can be expressed as:

[0027]

[0028] Where c is the number of categories, y ic is the true label of the i-th sample in category c, f(x i ;θ) c is the probability that the model predicts the i-th sample as category c;

[0029] The model parameters are updated by the gradient descent algorithm. In each iteration t, the model parameter update formula is

[0030] θ t+1 =θ t -η▽L(θ t )

[0031] Among them, ▽L(θ t) is the loss function L(θ) at t iterations θ t Gradient at the parameter;

[0032] Step 2.3: Model parameter upload and aggregation: Upload the updated local model parameters to the base station via satellite communication or ground wireless communication. The base station uses edge federated learning technology to aggregate the model parameters and obtain a new version of the global model.

[0033] Suppose there are m edge nodes, the updated model parameters of the jth edge node are The global model parameters are aggregated by weighted average aggregation, and the aggregated model parameters are θ new

[0034]

[0035] Among them, w j Aggregation weight

[0036]

[0037] Among them, n j is the number of local data samples of the jth edge node, n k is the number of local data samples of the kth edge node. During the transmission of model parameters, the transmission rate of model parameters must be less than or equal to the channel capacity to ensure reliable transmission.

[0038] As a further improvement of the present invention, the process of transmitting the model parameters based on air calculation in step 3 can be expressed as follows:

[0039] Step 3.1: Multiple devices transmit simultaneously through non-orthogonal transmission technology. The devices encode and map the information to be transmitted into codewords and then send them together. After receiving the superimposed signal, the receiving end detects and recovers the information. The signal received by the receiving end is:

[0040]

[0041] Where U is the number of devices, h u is the channel fading coefficient, x u is the codeword of device u, n is the noise; construct the objective function:

[0042] J=αU-βE

[0043] Among them, α and β are trade-off coefficients, U is the number of devices, and E is the transmission error measured by mean square error. Under the power constraint and channel capacity constraint conditions, the optimal device selection and transceiver parameter scheme are searched;

[0044] Step 3.2: Using the superposition characteristics of wireless channels, the devices simultaneously send signals that are superimposed in the air. The receiving end directly calculates the superimposed signals and ensures the synchronization of device signals.

[0045] As a further improvement of the present invention, the process of collaboration between the perception and computing tasks in step 4 can be expressed as follows:

[0046] Step 4.1: Calculate the time required for local processing of the perception task based on complexity and latency offload

[0047]

[0048] Among them, i is the perception task, C i is the computational complexity, is the computational rate of edge nodes processing sensing tasks; if If the latency is greater than the task delay requirement, the task is offloaded to the base station or edge server for calculation, meeting the latency requirement and reducing the local burden.

[0049] Step 4.2, consider energy offloading. If the total energy consumption of local processing of edge node processing tasks is greater than the energy consumption of remote node processing, the task is offloaded to the base station and edge server for calculation;

[0050] Step 4.3, multi-dimensional computing resource collaboration, when performing N-dimensional computing tasks, dimension n requires computing resources, and the total computing resources of the system are C total When:

[0051]

[0052] By calculating the task priority p n To allocate processor core resources, the allocation ratio n for:

[0053]

[0054] The number of cores that can be allocated in dimension n is:

[0055]

[0056] in, It is the number of cores of the system's total computing resources, which enables multi-dimensional computing collaboration, improves resource utilization and supports communication collaboration.

[0057] As a further improvement of the present invention, the process of function multiplexing and interference coordination in step 5 can be expressed as follows:

[0058] Step 5.1, design the joint beamforming target, set the weighted total performance and power minimization as the set goal, and use the communication channel capacity C, perception accuracy E s, calculation efficiency R c To calculate the total system performance P:

[0059] P=w1C-w2E s +w3R c

[0060] Among them, w1, w2, w3 are weighting coefficients;

[0061] Step 5.2, transmit beamforming, and calculate the optimal beamforming vector by solving the optimization problem using the Lagrange multiplier method. The optimization problem is:

[0062]

[0063] Where w is the beamforming vector and N is the dimension of the transmitted signal vector;

[0064] Step 5.3, receive beamforming, the processed signal is:

[0065] y processed =v H y

[0066] Among them, y processed is the signal after receiving beamforming, v is the receiving beamforming vector, H is the channel matrix, and y is the signal received by the receiving end;

[0067] Step 5.4, calculate the interference between communication and perception, communication and computation, and perception and computation respectively: aP c , Where a is the frequency domain overlap ratio, P c is the communication signal power, R t is the communication data transmission rate, B c is the computing task bandwidth, R s is the rate at which sensor data is generated, R proc It is the processing power of computing resources;

[0068] Step 5.5, beamforming coordination, is achieved by minimizing and in is the receive beamforming vector of the channel matrix H sensing device, is the receiving beamforming vector of the channel matrix H computing device, and the interference of communication on the perception and computing direction is reduced by designing the beamforming vector;

[0069] Step 5.6, resource allocation coordination, the total frequency band on the spectrum is B total , by reasonably allocating to communication B c , Perception B s , calculate the time slice ratio B calc; The total time is T total , allocate communication B c , Perception B s , calculate the time slice ratio B calc ; Use beamforming in space to separate signals and reduce interference.

[0070] The invention provides a satellite communication and edge network intelligent perception method based on federated learning, which has beneficial effects

[0071] 1. The present invention realizes the comprehensive collection of environmental information, target status and communication data through multi-source data collection and preprocessing, and uses satellite technology to expand the scope of data collection, ensuring the quality and availability of data.

[0072] 2. The present invention makes full use of distributed computing resources through edge federated learning model training and air computing technology, improves the efficiency and accuracy of model training, realizes continuous optimization of the global model, and reduces the model parameter transmission error.

[0073] 3. The present invention can flexibly offload part of the perception tasks to the base station, edge server or satellite for calculation, thereby improving resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is a flow chart of the present invention;

[0075] Figure 2 This is a diagram of software transmission, distribution and preprocessing of the present invention. DETAILED DESCRIPTION

[0076] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0077] The present invention proposes a satellite communication and edge network intelligent perception method based on federated learning. The invention flow chart is as follows: Figure 1 As shown, the steps of the present invention are described in detail below.

[0078] Step 1: Multi-source data collection and preprocessing: deploy wireless sensors and cameras to cover the edge of the network, collect environmental information, target status and communication data, use low-orbit satellites or geostationary satellites to collect perception data in areas that cannot be covered by the ground, and preprocess the data;

[0079] Step 1.1: Ground sensing data collection: Deploy wireless sensors and cameras at the edge of the network. The sensors fully cover the target area through the layout of topological structure planning to obtain the most complete environmental information, target status and communication data;

[0080] Step 1.2, satellite sensing data collection, uses low-orbit satellites and geostationary satellites to obtain sensing data of areas that cannot be covered by the ground. Low-orbit satellites achieve continuous coverage through a network of multiple satellites. The relationship between the orbit period and orbit altitude of a single low-orbit satellite can be expressed as:

[0081]

[0082] Among them, T LEO is the orbital period of the satellite, h LEO is the track height, R E is the radius of the Earth, G is the gravitational constant, M E is the mass of the Earth;

[0083] Step 1.3, using Huffman coding compression for character data, so that characters with high frequency of occurrence are represented by shorter codes, and characters with low frequency of occurrence are represented by longer codes, thereby achieving character data compression;

[0084] Step 1.4, using the LZ77 algorithm for text data, by setting the sliding window and the look-ahead window, looking for the longest string in the look-ahead window that matches the string in the sliding window, if a matching string is found, using a pointer and length to represent the matching string;

[0085] Step 1.5: Use JPEG compression on the image data. First, divide the image into 8×8 blocks. Let the image be I(x, y). Each block can be represented by I i,j (x, y), i and j represent the position of the small block in the image, x and y represent the image coordinates, and a discrete cosine transform is performed on each small block:

[0086]

[0087] Among them, F(u,v) is the coefficient after discrete cosine transform, u and v are the frequency coordinates after transform; then the discrete cosine transform coefficient is quantized, and the quantized coefficient F q (u,v) can be expressed as:

[0088]

[0089] Among them, Q(u,v) is the quantization matrix, and finally the quantized coefficients are entropy encoded and encoded into compressed data.

[0090] Step 2: Edge federated learning model training: fine-tune the pre-trained global model using the local data set at the edge node, update the local model parameters, and upload the updated local model parameters to the base station via satellite communication or ground wireless communication. The base station uses edge federated learning technology to aggregate the model parameters and obtain a new version of the global model.

[0091] In the satellite-ground fusion network environment, the edge federated learning model is mainly used to realize distributed machine learning across regions and devices in the satellite-ground fusion network environment, while protecting the privacy and security of user data and optimizing resource utilization and system performance.

[0092] Step 2.1: Local model training. Let the pre-trained global model be f(x;θ), where x is the input data and θ is the model parameter, and establish a local dataset. Where n is the number of local data samples, x i The i-th input sample, y i is the corresponding label;

[0093] Step 2.2, update the model parameters and use the cross entropy loss function to train the model. The cross entropy loss function can be expressed as:

[0094]

[0095] Where c is the number of categories, y ic is the true label of the i-th sample in category c, f(x i ;θ) c is the probability that the model predicts the i-th sample as category c;

[0096] The model parameters are updated by the gradient descent algorithm. In each iteration t, the model parameter update formula is

[0097]

[0098] in, is the loss function L(θ) at t iterations θ t Gradient at the parameter;

[0099] Step 2.3: Model parameter upload and aggregation: Upload the updated local model parameters to the base station via satellite communication or ground wireless communication. The base station uses edge federated learning technology to aggregate the model parameters and obtain a new version of the global model.

[0100] Suppose there are m edge nodes, the updated model parameters of the jth edge node are The global model parameters are aggregated by weighted average aggregation, and the aggregated model parameters are θ new

[0101]

[0102] Among them, w j Aggregation weight

[0103]

[0104] Among them, nj is the number of local data samples of the jth edge node, n k is the number of local data samples of the kth edge node. During the transmission of model parameters, the transmission rate of model parameters must be less than or equal to the channel capacity to ensure reliable transmission.

[0105] Step 3: Model parameter transmission based on air computing, using non-orthogonal transmission technology and air computing technology, design joint device selection and transceiver schemes to maximize the number of devices participating in federated learning and minimize model parameter transmission errors;

[0106] Step 3.1: Multiple devices transmit simultaneously through non-orthogonal transmission technology. The devices encode and map the information to be transmitted into codewords and then send them together. After receiving the superimposed signal, the receiving end detects and recovers the information. The signal received by the receiving end is:

[0107]

[0108] Where U is the number of devices, h u is the channel fading coefficient, x u is the codeword of device u, n is the noise; construct the objective function:

[0109] J=αU-βE

[0110] Among them, α and β are trade-off coefficients, U is the number of devices, and E is the transmission error measured by mean square error. Under the power constraint and channel capacity constraint conditions, the optimal device selection and transceiver parameter scheme are searched;

[0111] Step 3.2: Using the superposition characteristics of wireless channels, the devices simultaneously send signals that are superimposed in the air. The receiving end directly calculates the superimposed signals and ensures the synchronization of device signals.

[0112] Step 4: Coordinate perception and computing tasks. According to the complexity and latency requirements of the perception tasks, some perception tasks are offloaded to base stations, edge servers or satellites for computing, reducing the computing burden of edge nodes and ground networks, achieving multi-dimensional computing on the same resource block, improving resource utilization, and supporting the coordination of satellite communications and ground communications. The software transmission distribution and preprocessing diagram is shown in the figure. Figure 2 As shown;

[0113] Step 4.1: Calculate the time required for local processing of the perception task based on complexity and latency offload

[0114]

[0115] Among them, i is the perception task, C i is the computational complexity, is the computational rate of edge nodes processing sensing tasks; if If the latency is greater than the task delay requirement, the task is offloaded to the base station or edge server for calculation, meeting the latency requirement and reducing the local burden.

[0116] Step 4.2, consider energy offloading. If the total energy consumption of local processing of edge node processing tasks is greater than the energy consumption of remote node processing, the task is offloaded to the base station and edge server for calculation;

[0117] Step 4.3, multi-dimensional computing resource collaboration, when performing N-dimensional computing tasks, dimension n requires computing resources, and the total computing resources of the system are C total When:

[0118]

[0119] By calculating the task priority p n To allocate processor core resources, the allocation ratio n for:

[0120]

[0121] The number of cores that can be allocated in dimension n is:

[0122]

[0123] in, It is the number of cores of the system's total computing resources, which enables multi-dimensional computing collaboration, improves resource utilization and supports communication collaboration.

[0124] Step 5: Functional reuse and interference coordination: design a joint transmit and receive beamforming algorithm based on maximizing weighted total performance and minimizing total transmit power, coordinate interference between communication, perception, and computing, improve overall system performance, and support coordination between satellite and terrestrial communications.

[0125] Step 5.1, design the joint beamforming target, set the weighted total performance and power minimization as the set goal, and use the communication channel capacity C, perception accuracy E s , calculation efficiency R c To calculate the total system performance P:

[0126] P=w1C-w2E s +w3R c

[0127] Among them, w1, w2, w3 are weighting coefficients;

[0128] Step 5.2, transmit beamforming, and calculate the optimal beamforming vector by solving the optimization problem using the Lagrange multiplier method. The optimization problem is:

[0129]

[0130] Where w is the beamforming vector and N is the dimension of the transmitted signal vector;

[0131] Step 5.3, receive beamforming, the processed signal is:

[0132] y processed =v H y

[0133] Among them, y processed is the signal after receiving beamforming, v is the receiving beamforming vector, H is the channel matrix, and y is the signal received by the receiving end;

[0134] Step 5.4, calculate the interference between communication and perception, communication and computation, and perception and computation respectively: aP c , Where a is the frequency domain overlap ratio, P c is the communication signal power, R t is the communication data transmission rate, B c is the computing task bandwidth, R s is the rate at which sensor data is generated, R proc It is the processing power of computing resources;

[0135] Step 5.5, beamforming coordination, is achieved by minimizing and in is the receive beamforming vector of the channel matrix H sensing device, is the receiving beamforming vector of the channel matrix H computing device, and the interference of communication on the perception and computing direction is reduced by designing the beamforming vector;

[0136] Step 5.6, resource allocation coordination, the total frequency band on the spectrum is B total , by reasonably allocating to communication B c , Perception B s , calculate the time slice ratio B calc ; The total time is T total , allocate communication B c , Perception B s , calculate the time slice ratio B calc ; Use beamforming in space to separate signals and reduce interference.

[0137] Step 6: Intelligent business orchestration and optimization. Based on the global network perception and business characteristics, the edge control center performs intelligent business orchestration. According to the needs of different application scenarios, it dynamically adjusts the allocation and scheduling of communication, perception and computing resources, and coordinates the collaborative work of the ground network and the satellite network.

[0138] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A satellite communication and edge network intelligent perception method based on federated learning, the specific steps are as follows, and the characteristics are as follows: Step 1: Multi-source data collection and preprocessing: deploy wireless sensors and cameras to cover the edge of the network, collect environmental information, target status and communication data, use low-orbit satellites or geostationary satellites to collect perception data in areas that cannot be covered by the ground, and preprocess the data; Step 2: Edge federated learning model training: fine-tune the pre-trained global model using the local data set at the edge node, update the local model parameters, and upload the updated local model parameters to the base station via satellite communication or ground wireless communication. The base station uses edge federated learning technology to aggregate the model parameters and obtain a new version of the global model. Step 3: Model parameter transmission based on air computing, using non-orthogonal transmission technology and air computing technology, design joint device selection and transceiver schemes to maximize the number of devices participating in federated learning and minimize model parameter transmission errors; Step 4: Coordinate the perception and computing tasks. According to the complexity and latency requirements of the perception tasks, some perception tasks are offloaded to base stations, edge servers or satellites for computing, reducing the computing burden of edge nodes and ground networks, achieving multi-dimensional computing on the same resource block, improving resource utilization, and supporting the coordination of satellite communications and ground communications. Step 5: Functional reuse and interference coordination: design a joint transmit and receive beamforming algorithm based on maximizing weighted total performance and minimizing total transmit power, coordinate interference between communication, perception, and computing, improve overall system performance, and support coordination between satellite and terrestrial communications. Step 6: Intelligent business orchestration and optimization. Based on the global network perception and business characteristics, the edge control center performs intelligent business orchestration. According to the needs of different application scenarios, it dynamically adjusts the allocation and scheduling of communication, perception and computing resources, and coordinates the collaborative work of the ground network and the satellite network.

2. The satellite communication and edge network intelligent perception method based on federated learning according to claim 1, characterized in that: The multi-source data collection and preprocessing in step 1 can be expressed as: Step 1.1: Ground sensing data collection: wireless sensors and cameras are deployed at the edge of the network. The sensors fully cover the target area through the layout of topological structure planning to obtain the most complete environmental information, target status and communication data; Step 1.2, satellite sensing data collection, uses low-orbit satellites and geostationary satellites to obtain sensing data of areas that cannot be covered by the ground. Low-orbit satellites achieve continuous coverage through a network of multiple satellites. The relationship between the orbit period and orbit altitude of a single low-orbit satellite can be expressed as: Among them, T LEO is the orbital period of the satellite, h LEO is the track height, R E is the radius of the Earth, G is the gravitational constant, M E is the mass of the Earth; Step 1.3, using Huffman coding compression for character data, so that characters with high frequency of occurrence are represented by shorter codes, and characters with low frequency of occurrence are represented by longer codes, thereby achieving character data compression; Step 1.4, using the LZ77 algorithm for text data, by setting the sliding window and the look-ahead window, looking for the longest string in the look-ahead window that matches the string in the sliding window, if a matching string is found, using a pointer and length to represent the matching string; Step 1.5: Use JPEG compression on the image data. First, divide the image into 8×8 blocks. Let the image be I(x, y). Each block can be represented by I i,j (x, y), i and j represent the position of the small block in the image, x and y represent the image coordinates, and a discrete cosine transform is performed on each small block: Among them, F(u,v) is the coefficient after discrete cosine transform, u and v are the frequency coordinates after transform; then the discrete cosine transform coefficient is quantized, and the quantized coefficient F q (u,v) can be expressed as: Among them, Q(u,v) is the quantization matrix, and finally the quantized coefficients are entropy encoded and encoded into compressed data.

3. The satellite communication and edge network intelligent perception method based on federated learning according to claim 1, characterized in that: The process of edge federated learning model training in step 2 can be expressed as follows: Step 2.1: Local model training. Let the pre-trained global model be f(x;θ), where x is the input data and θ is the model parameter, and establish a local dataset. Where n is the number of local data samples, x i The i-th input sample, y i is the corresponding label; Step 2.2, update the model parameters and use the cross entropy loss function to train the model. The cross entropy loss function can be expressed as: Where c is the number of categories, y ic is the true label of the i-th sample in category c, f(x i ;θ) c is the probability that the model predicts the i-th sample as category c; The model parameters are updated by the gradient descent algorithm. In each iteration t, the model parameter update formula is in, is the loss function L(θ) at t iterations θ t Gradient at the parameter; Step 2.3: Model parameter upload and aggregation: Upload the updated local model parameters to the base station via satellite communication or ground wireless communication. The base station uses edge federated learning technology to aggregate the model parameters and obtain a new version of the global model. Suppose there are m edge nodes, the updated model parameters of the jth edge node are The global model parameters are aggregated by weighted average aggregation, and the aggregated model parameters are θ new Among them, w j Aggregation weight Among them, n j is the number of local data samples of the jth edge node, n k is the number of local data samples of the kth edge node. During the transmission of model parameters, the transmission rate of model parameters must be less than or equal to the channel capacity to ensure reliable transmission.

4. The satellite communication and edge network intelligent perception method based on federated learning according to claim 1, characterized in that: The process of model parameter transmission based on air calculation in step 3 can be expressed as follows: Step 3.1: Multiple devices transmit simultaneously through non-orthogonal transmission technology. The devices encode and map the information to be transmitted into codewords and then send them together. After receiving the superimposed signal, the receiving end detects and recovers the information. The signal received by the receiving end is: Where U is the number of devices, h u is the channel fading coefficient, x u is the codeword of device u, n is the noise; construct the objective function: J=αU-βE Among them, α and β are trade-off coefficients, U is the number of devices, and E is the transmission error measured by mean square error. Under the power constraint and channel capacity constraint conditions, the optimal device selection and transceiver parameter scheme are searched; In step 3.2, using the superposition characteristics of wireless channels, the devices simultaneously send signals to superimpose in the air, and the receiving end directly calculates the superimposed signals while ensuring device signal synchronization.

5. The satellite communication and edge network intelligent perception method based on federated learning according to claim 1, characterized in that: The process of coordinating the perception and computation tasks in step 4 can be expressed as follows: Step 4.1: Calculate the time required for local processing of the perception task based on complexity and latency offload Among them, i is the perception task, C i is the computational complexity, is the computational rate of edge nodes processing sensing tasks; if If the latency is greater than the task delay requirement, the task is offloaded to the base station or edge server for calculation, meeting the latency requirement and reducing the local burden. Step 4.2, consider energy offloading. If the total energy consumption of local processing of edge node processing tasks is greater than the energy consumption of remote node processing, the task is offloaded to the base station and edge server for calculation; Step 4.3, multi-dimensional computing resource collaboration, when performing N-dimensional computing tasks, dimension n requires computing resources, and the total computing resources of the system are C total When: By calculating the task priority p n To allocate processor core resources, the allocation ratio n for: The number of cores that can be allocated in dimension n is: in, It is the number of cores of the system's total computing resources, which enables multi-dimensional computing collaboration, improves resource utilization and supports communication collaboration.

6. The satellite communication and edge network intelligent perception method based on federated learning according to claim 1, characterized in that: The process of functional reuse and interference coordination in step 5 can be expressed as follows: Step 5.1, design the joint beamforming target, set the weighted total performance and power minimization as the set goal, and use the communication channel capacity C, perception accuracy E s , calculation efficiency R c To calculate the total system performance P: P=w1C-w2E s +w3R c Among them, w1, w2, w3 are weighting coefficients; Step 5.2, transmit beamforming, and calculate the optimal beamforming vector by solving the optimization problem using the Lagrange multiplier method. The optimization problem is: Where w is the beamforming vector and N is the dimension of the transmitted signal vector; Step 5.3, receive beamforming, the processed signal is: and processed =v H and Among them, y processed is the signal after receiving beamforming, v is the receiving beamforming vector, H is the channel matrix, and y is the signal received by the receiving end; Step 5.4, calculate the interference between communication and perception, communication and computation, and perception and computation respectively: aP c , Where a is the frequency domain overlap ratio, P c is the communication signal power, R t is the communication data transmission rate, B c is the computing task bandwidth, R s is the rate at which sensor data is generated, R proc It is the processing power of computing resources; Step 5.5, beamforming coordination, is achieved by minimizing and in is the receive beamforming vector of the channel matrix H sensing device, is the receiving beamforming vector of the channel matrix H computing device, and the interference of communication on the perception and computing direction is reduced by designing the beamforming vector; Step 5.6, resource allocation coordination, the total frequency band on the spectrum is B total , by reasonably allocating to communication B c , Perception B s , calculate the time slice ratio B calc ; The total time is T total , allocate communication B c , Perception B s , calculate the time slice ratio B calc ; Use beamforming in space to separate signals and reduce interference.

Images