An aerial federated learning system integrating synaesthesia and computing and its learning method

Abstract

The present invention belongs to the field of mobile communication technology and discloses an airborne federated learning system and a learning method thereof that integrates synaesthesia and computing. The learning system includes a model learning module, a device perception module, a mobile agent, and a core server. The learning method is as follows: the mobile agent first trains a local model based on regional representative weights and then uploads the local model. The core server perceives the location and status information of the mobile agent, updates the regional representative weights, and sends the corresponding power compensation coefficients to all mobile agents. The core server then aggregates the global model and broadcasts the global model for the next round of federated learning. In non-independent and identically distributed scenarios, the present invention considers the location information and channel status of all mobile agents and updates the global model based on local gradient deviation to reduce the impact of local sample deviation.

Description

Technical Field

[0001] The present invention belongs to the field of mobile communication technology, and more specifically relates to an airborne federated learning system integrating synaesthesia and computing and a learning method thereof. Background Art

[0002] In today's rapidly developing digital world, with the increasing communication computing and caching capabilities of various mobile intelligent agents (MAs) (such as mobile base stations, unmanned vehicles, mobile phones, smart gateways, etc.), the future wireless communication paradigm is undergoing a paradigm shift from the Internet of Everything to the Internet of Intelligence. This is expected to support the development of various artificial intelligence applications and services. For example, augmented reality and virtual reality are gradually becoming part of people's lives. At the same time, intelligent large models are gradually being valued by researchers. However, training intelligent large models requires collecting a large amount of data, which causes the network data traffic to grow exponentially. This trend poses a huge challenge to the efficiency and stability of communication networks, especially in the current environment with extremely strict requirements for low latency and privacy protection. Traditional centralized computing frameworks have gradually failed to meet the development needs of modern Internet of Things (IoT) devices and applications due to their limitations in processing large-scale real-time data. As a typical FL algorithm, average federated learning (FedAvg) iteratively executes two stages: (1) MA receives the global model from the core server and updates the multi-step local model; (2) MA sends the updated local model to the core server for aggregation to obtain a new global model. As the computing power of MA increases, the practicality and applicability of this model are also expanding. While federated learning (FL) improves resource constraints and protects data privacy, transmitting these local models over wireless networks still requires significant resources, which has become a bottleneck for implementing FL frameworks in real-world wireless networks. Therefore, much room for optimization remains in wireless federated learning (WFL). Recent research has considered asynchronous mechanisms, user scheduling, bandwidth-aware adaptation, and gradient quantization. However, these approaches ignore issues at the physical and network layers, focusing solely on communication and computation, without considering the integration of communication and computation.

[0003] In the second phase of the average federated learning (FedAvg) iteration execution, traditional multiple access schemes, such as orthogonal frequency division multiple access, will cause wireless resource consumption to increase linearly with the increase in the number of user access and the reduction in communication bandwidth, resulting in a decrease in the efficiency of wireless federated learning (WFL) aggregation and occupying a large amount of communication resources.

[0004] Most existing studies do not consider the data heterogeneity caused by the data that needs to be collected to train large intelligent models. This results in local models generally being geographically regional. Data labels within a region are biased. This is because the behavior of devices in different regions is different. Regions with large geographical distance differences may have different local model weights. However, existing studies do not combine air computing and device perception. They ignore the hidden features in the superimposed signals during the over the air computing (OAC) process. The arrival angle estimator based on deep learning in the existing technology is good at directly extracting highly abstract and representative features from large amounts of data, which enables it to autonomously learn the effects of channel state differences and low signal-to-noise ratio. However, this study simply perceived the location of the device, but did not combine the communication and computing functions. Summary of the Invention

[0005] To address the aforementioned technical issues, the present invention provides an integrated over-the-air federated learning system and learning method. To address local model gradient bias caused by regional datasets, the present invention proposes a new region-selective over-the-air federated learning algorithm. This algorithm considers the location and channel state information of all mobile agents to update the regional weight matrix within the core server's service area. Furthermore, to separately obtain the location and channel state information of all mobile agents, the present invention uses a neural network group to perceive MAs.

[0006] In order to achieve the above object, the present invention is achieved through the following technical solutions:

[0007] The present invention is an airborne federated learning system integrating synaesthesia and computing, and the airborne federated learning system includes:

[0008] Model learning module: consists of the air federated learning method for region selection (Air-FedRS);

[0009] Device perception module: The device perception module is an NNG user sensor, which processes the superimposed signal to perceive the location and signal status of the mobile agent that uploaded the local model;

[0010] Mobile Agent (MA): collects data or collects data sent by personal devices in different regions, trains local models based on regional representative weights, and uploads local models to the core server after training to aggregate the global model;

[0011] The core server (CS) senses the location and status of mobile agents, updates regional representative weights, and sends corresponding power compensation coefficients to all N mobile agents. After receiving all local models, the core server aggregates and broadcasts the global model, inputs it into the model learning module, and uses the regional selection air federated learning algorithm (Air-FedRS) for the next round of federated learning.

[0012] A further improvement of the present invention is that the NNG user sensor includes a classifier and multiple regression heads, the classifier is an agent number sensor, the regression head is a single-agent sensor, a dual-agent sensor, and a multi-agent sensor, the agent number sensor is used to estimate the number of all mobile agents of the superimposed signal, the single-agent sensor, the dual-agent sensor, and the multi-agent sensor are used to perceive the position and channel state information of all different mobile agents, and the NNG user sensor utilizes the representative features of the superimposed signal in the air federated learning to perceive the position of different mobile agents and their respective channel states, thereby supplementing the perception function of the traditional air federated learning framework.

[0013] The NNG user sensor processes the superimposed signals to perceive the location and signal status of the local model mobile agent. The specific steps include:

[0014] Receive signal x

[0015]

[0016] Calculate the sampling covariance matrix R:

[0017]

[0018] Among them, V is the sampling window size, x H is the conjugate transpose of the received signal, L is the number of uploaded signals, l is the uploaded signal index, a is the channel gain, is the direction angle, n is the noise;

[0019] The sampling covariance matrix R is input into the NNG user sensor for perception. The sampling covariance matrix R first enters the agent number sensor to obtain the number of all mobile agents N. t Then, the sampling covariance matrix R is input into the regression head to obtain the positions and channel states of all different mobile agents. If N t =1, then use a single agent sensor; N t =2, then use the dual agent sensor. If N t =3, then use three-agent perceptron, if N t =X, then use X as the proxy perceptron, and so on.

[0020] A further improvement of the present invention is that the air federated learning method (Air-FedRS) specifically includes the following steps:

[0021] Step 1: Initialize the service range weight matrix M according to the service range and specific requirements of the core server, initialize the global model and broadcast it to all mobile agents participating in the aerial federated learning, and determine the total loss function;

[0022] In step 2, all mobile agents in step 1 receive the global model and perform local model training based on the service range weight matrix M. After completing the local model training, they upload the test signal through the over-the-air calculation algorithm to obtain the power compensation coefficient. After preprocessing and encoding the local model, they synchronize the time and frequency by uploading the local model through the over-the-air calculation algorithm and aggregating the global model.

[0023] Step 3: After obtaining the global model in step 2, the core server calculates the global loss, updates the service range weight matrix M based on the global loss and the locations of all mobile agents, and finally broadcasts the global model.

[0024] A further improvement of the present invention is that step 1 specifically includes the following steps:

[0025] Step 1.1, determine all mobile agents within the service range of the core server and determine the regional division according to the pitch angle, that is, divide the core servers in different regions according to their degree;

[0026] Step 1.2: Initialize the service range weight matrix M: Define the regional representative weight of the nth mobile agent updated in the tth round as Where M(ρ n ,θ n ) is the core server corresponding to the location information (ρ n ,θ n ) is the service range weight matrix element, M is the service range weight matrix, in order to limit the regional representative weight to expand the occasional regional data set to cause the impact of the model accuracy. The range is (1, +∞), and the default step size is 10°. Therefore, the size of the M matrix is ​​[36, 10]. Before starting the first round of aggregation, all elements of the service range weight matrix M are initialized to 1;

[0027] Step 1.3. Determine the global loss function: global loss function for global model training Defined as

[0028]

[0029] Among them, N is the total number of mobile agents, N t is the number of mobile agents that upload local models in the tth round of aggregation, Nt <N, is the set of communication rounds, T is the total number of communications, f n (w n ) is the loss function of the nth mobile agent:

[0030]

[0031] Where loss(x n,i ,y n,i ;w n ) is for a given global model parameter w n Under the premise that the nth mobile agent has n,i ,y n,i ), is the training set of each n-th mobile agent, x n is the input data accessible to the nth mobile agent, y n For the corresponding label is the collection of all inputs, A collection of all tags.

[0032] A further improvement of the present invention is that step 2 specifically includes the following steps:

[0033] Step 2.1: All mobile agents are based on the loss function f n (w n ) Perform local model training;

[0034] Step 2.2: After all mobile agents complete local model training, they use the air computing algorithm to aggregate local models. The global model after the t-th round of aggregation is w t for:

[0035]

[0036] Among them, μ t is the learning rate, A t is the total number of mobile agents selected for transmission in round t, is the label chosen by the nth mobile agent in the tth round of communication, when σ is a given threshold; is the power compensation coefficient, is the denoising coefficient of the core server, is the channel gain of the nth mobile agent in the tth communication round, for The Hessian matrix of ; the coefficients need to take into account channel inversion and power alignment to satisfy the paradigm of air calculation. The optimal solution is:

[0037]

[0038] Step 2.3: The core server receives the signal of the global model, which is expressed as

[0039]

[0040] in, is the local model of the nth mobile agent after the tth round of training, P n,t is the maximum transmission power of the nth mobile agent, is the post-processing function;

[0041] The information the core server expects to receive is:

[0042]

[0043] Among them, f n (w n ) is the loss function of the nth mobile agent;

[0044] Then the output of the integrated aerial federated learning system on the core server side is

[0045]

[0046] Where: s t Receives signals from the global model for the core server.

[0047] A further improvement of the present invention is that step 3 specifically includes the following steps:

[0048] Step 3.1: The core server calculates the global loss function

[0049] Step 3.2: Service range weight matrix M. After each round of aggregation, if the new global model loss F t (w) is less than or equal to the global model loss F of the previous round t-1 (w), then give the updated position information (ρ n ,θ n ) corresponds to the elements in the service range weight matrix M multiplied by a reward coefficient α. After each round of aggregation, if the new global model loss F t (w) is greater than the global model loss F of the previous round t-1 (w), then give the updated position information (ρ n ,θn ) corresponds to the element in M ​​multiplied by a penalty coefficient β, that is

[0050]

[0051] Among them, α<1,β>1.

[0052] The present invention also provides an airborne federated learning method integrating synaesthesia and computing, which specifically includes the following steps:

[0053] Step (1), the core server initializes the global model parameters w 0 , service range weight matrix M, and then the global model parameters w 0 And the service range weight matrix M is broadcast to all mobile agents;

[0054] Step (2): All mobile agents receive the global model parameter w 0 Then select the local model to train and get the local model

[0055] Step (3): All mobile agents upload test signals to the core server, and the NNG sensor senses the user's channel status information, and then calculates the power compensation coefficient.

[0056] Step (4), local model preprocessing and encoding and time-frequency synchronization are performed, and the local model is uploaded through the air computing algorithm to aggregate the global model w t ;

[0057] Step (5): The core server receives the superimposed signal, decodes it, and post-processes it to obtain the global model w t ,At the same time, the positions of all mobile agents are obtained by calculating the sampling covariance matrix R of the superimposed signal and using the NNG perceptron;

[0058] Step (6): Calculate the global loss function through the global model And update the service range weight matrix M according to the locations of all mobile agents that uploaded the local model;

[0059] Step (7), the core server broadcasts the new global model w * .

[0060] The beneficial effects of the present invention are as follows: the present invention proposes a new integrated inter-sensory computing air federated learning system (ISCC Air FedRS Framework) and performs convergence analysis on it.

[0061] The present invention proposes an air federated learning algorithm with regional selection (Air-FedRS) to solve the impact of regional datasets on the effect of air federated learning.

[0062] This invention combines the characteristics of over-the-air federated learning and innovatively uses superimposed signals to perceive the location and channel status of uploaded local model agents.

[0063] The present invention proposes a closed-form solution for the optimal denoising coefficient. Based on this, the present invention studies the training of a neural network group to perceive information and assist the training of an airborne federated learning system.

[0064] Because the framework adopts a modular design, when new technologies and algorithms emerge in the future, the modules in the framework can be replaced. This greatly improves the flexibility of the framework and provides room for future optimization. This shows that the framework has a bright future and a wide range of applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a model diagram of the aerial federated learning system that integrates synaesthesia and computing in the present invention.

[0066] Figure 2 This is a technical block diagram of the in-flight federated learning system that integrates synergy and computing.

[0067] Figure 3 It is a technical block diagram of the NNG sensor of the present invention.

[0068] Figure 4 This is a rendering of the present invention based on the NNG sensor.

[0069] Figure 5 This is a diagram showing the optimization effect of the integrated aerial federated learning system of the present invention.

[0070] Figure 6 This is a diagram showing the impact of hyperparameters in the integrated aerial federated learning system of synergy computing of the present invention. DETAILED DESCRIPTION

[0071] The following diagrams illustrate embodiments of the present invention. For clarity, many practical details are included in the following description. However, it should be understood that these practical details are not intended to limit the present invention. In other words, in some embodiments of the present invention, these practical details are not essential.

[0072] like Figure 1-2 As shown, the present invention provides an airborne federated learning system integrating synergy and computing, and the airborne federated learning system integrating synergy and computing includes:

[0073] Model learning module: consists of the air federated learning method for region selection (Air-FedRS);

[0074] Device perception module: The device perception module is an NNG user sensor, which processes superimposed signals to perceive the location and signal status of the uploaded local model mobile agent; the NNG user sensor includes a classifier and multiple regression heads, the classifier is an agent quantity sensor, and the regression head is a single agent sensor, a dual agent sensor, and a multi-agent sensor. The agent quantity sensor is used to estimate the number of all mobile agents in the superimposed signal, and the single agent sensor, the dual agent sensor, and the multi-agent sensor are used to perceive the location and channel status information of all different mobile agents. The NNG user sensor uses the representative features of the superimposed signals in air federated learning to perceive the location and respective channel status of different mobile agents, supplementing the perception function of the traditional air federated learning framework. The NNG user sensor processes superimposed signals to perceive the location and signal status of the local model mobile agent, specifically including the following steps:

[0075] Receive signal x

[0076]

[0077] Calculate the sampling covariance matrix R:

[0078]

[0079] Among them, V is the sampling window size, x H is the conjugate transpose of the received signal, L is the number of uploaded signals, l is the uploaded signal index, a is the channel gain, (θ l ) is the direction angle, n is the noise;

[0080] The sampling covariance matrix R is input into the NNG user sensor for perception. The sampling covariance matrix R first enters the agent number sensor to obtain the number of all mobile agents N. t Then, the sampling covariance matrix R is input into the regression head to obtain the positions and channel states of all different mobile agents. If N t =1, then use a single agent sensor; N t =2, then use the dual agent sensor. If N t =3, then use three-agent perceptron, and so on. If N t =X, then use X proxy perceptron.

[0081] Mobile Agent (MA): collects data or collects data sent by personal devices in different regions, trains local models based on regional representative weights, and uploads local models to the core server after training to aggregate the global model;

[0082] The core server (CS) senses the location and status of mobile agents, updates regional representative weights, and sends corresponding power compensation coefficients to all N mobile agents. After receiving all local models, the core server aggregates and broadcasts the global model, inputs it into the model learning module, and uses the regional selection air federated learning algorithm (Air-FedRS) for the next round of federated learning.

[0083] like Figure 1-2 As shown, the air federated learning method (Air-FedRS) proposed in the present invention specifically includes the following steps:

[0084] Step 1: Initialize the service range weight matrix M according to the service range and specific requirements of the core server, initialize the global model and broadcast it to all mobile agents participating in the aerial federated learning, and determine the total loss function;

[0085] In step 2, all mobile agents in step 1 receive the global model and perform local model training based on the service range weight matrix M. After completing the local model training, they upload the test signal through the over-the-air calculation algorithm to obtain the power compensation coefficient. After preprocessing and encoding the local model, they synchronize the time and frequency by uploading the local model through the over-the-air calculation algorithm and aggregating the global model.

[0086] Step 3: After obtaining the global model in step 2, the core server calculates the global loss, updates the service range weight matrix M based on the global loss and the locations of all mobile agents, and finally broadcasts the global model.

[0087] Based on this invention, an air federated learning method integrating synaesthesia and computing is proposed for an air federated learning system, which specifically includes the following steps:

[0088] Step (1), the core server initializes the global model parameters w 0 , service range weight matrix M, and then the global model parameters w 0 And the service range weight matrix M is broadcast to all mobile agents;

[0089] Step (2): All mobile agents receive the global model parameter w 0 Then select the local model to train and get the local model

[0090] Step (3): All mobile agents upload test signals to the core server, and the NNG sensor senses the user's channel status information, and then calculates the power compensation coefficient.

[0091] Step (4), local model preprocessing and encoding and time-frequency synchronization are performed, and the local model is uploaded through the air computing algorithm to aggregate the global model wt ;

[0092] Step (5): The core server receives the superimposed signal, decodes it, and post-processes it to obtain the global model w t ,At the same time, the positions of all mobile agents are obtained by calculating the sampling covariance matrix R of the superimposed signal and using the NNG perceptron;

[0093] Step (6): Calculate the global loss function through the global model And update the service range weight matrix M according to the locations of all mobile agents that uploaded the local model;

[0094] Step (7), the core server broadcasts the new global model w * .

[0095] The present invention is Figure 1 For this study, a federated learning scenario assisted by all mobile agents based on 5G or 6G wireless communications was proposed. A group of all mobile agents, such as unmanned vehicles and drones, is clustered around a core server. All mobile agents move irregularly within a fixed range but are relatively evenly distributed within the region. They collect data from neighboring terminals as datasets for local model training. Data sets collected from different regions are non-independent and identically distributed (Non-IID). This can affect the global gradient during the federated learning process, manifesting as a difference between local and global gradients. All mobile agents jointly train on the same task. They connect to the core server via backhaul links and collaborate to achieve federated learning. In each round of local federated learning, all mobile agents first collect data in different regions. Their geographic locations do not change significantly during each round of collection. Then, all mobile agents remain stationary and train their local models. Due to differences in the size and computing power of the collected datasets, and due to the strict synchronization of transmission by the senders using Open Access Calling (OAC), all mobile agents cannot simultaneously participate in each round of local model upload. This results in local models being biased towards regions and data. Local models are sent to the core server via OAC. The core server performs decoding analysis to obtain a global model, and then broadcasts the global model to all mobile agents and terminals.

[0096] The present invention models the problem as minimizing the error between the signal estimated by the core server and the error-free signal. Assume that there are N mobile agents in the area served by the core server who jointly train a model consisting of d parameters. These parameters can be represented by the vector w n express, Where n is the index of the nth mobile agent. All mobile agents can be accessed through index n. is the set of all mobile agents, Each n-th mobile agent has its own training set in x n is the input data accessible to the nth mobile agent, y n For the corresponding label. The collection of all inputs. A collection of all tags.

[0097] The present invention provides an air federated learning method (Air-FedRS), which specifically includes the following steps:

[0098] Step 1: Initialize the service range weight matrix M according to the service range and specific requirements of the core server, initialize the global model and broadcast it to all mobile agents participating in the aerial federated learning, and determine the total loss function.

[0099] In this step, the present invention considers a federated learning scenario based on MAs assistance under 5G or 6G wireless communication, such as Figure 1 As shown in the figure, a group of mobile agents, such as unmanned vehicles and drones, are clustered around a core server. These agents move randomly within a fixed range but are relatively evenly distributed within the region. They collect data from neighboring terminals as datasets for local model training. Data sets collected from different regions are non-independent and identically distributed (Non-IID). This can affect the global gradient during federated learning, manifesting as a difference between local and global gradients. All mobile agents train on the same task. They connect to the core server via backhaul links and collaborate to achieve federated learning. In each round of local federated learning, all mobile agents first collect data in different regions. Their geographic locations do not change significantly during each round of collection. Then, all mobile agents remain stationary and train their local models. Due to differences in the size and computing power of the collected datasets, and due to OAC's strict synchronization requirement for senders, all mobile agents cannot simultaneously participate in each round of local model upload. This can lead to local models favoring regions and data. Local models are sent to the core server via OAC. The core server decodes and analyzes the local model to generate a global model. This global model is then broadcast to all mobile agents and terminals.

[0100] The specific steps include:

[0101] Step 1.1: Determine all mobile agents within the service range of the core server and determine the area division based on the pitch angle;

[0102] Step 1.2: Initialize the service range weight matrix M: Define the regional representative weight of the nth mobile agent updated in the tth round as Where M(ρn ,θ n ) is the core server corresponding to the location information (ρ n ,θ n ) is a service range weight matrix element, M is a service range weight matrix, in order to limit the influence of regional representative weights on the accuracy of the model caused by the occasional regional data set. The range is (1, +∞), and the default step size is 10°. Therefore, the size of the M matrix is ​​[36, 10]. Before starting the first round of aggregation, all elements of the service range weight matrix M are initialized to 1;

[0103] Step 1.3. Determine the global loss function: global loss function for global model training Defined as

[0104]

[0105] Among them, N is the total number of mobile agents, N t is the number of mobile agents that upload local models in the tth round of aggregation, N t <N, is the set of communication rounds, T is the total number of communications, f n (w n ) is the loss function of the nth mobile agent:

[0106]

[0107] Where loss(x n,i ,y n,i ;w n ) is for a given global model parameter w n Under the premise that the nth mobile agent has n,i ,y n,i ), is the training set of each n-th mobile agent, x n is the input data accessible to the nth mobile agent, y n For the corresponding label is the collection of all inputs, A collection of all tags. Where n is the nth mobile agent index.

[0108] make is the local model of MA n after the tth round of training, that is

[0109]

[0110] Among them, μ t is the learning rate. Finally, the global model w after the t-th round of aggregation is obtained t for

[0111]

[0112] Define the global gradient:

[0113]

[0114] where δ n is the deviation between the local gradient and the global average gradient of MAn due to the influence of regional datasets.

[0115] In step 2, all mobile agents in step 1 receive the global model and perform local model training based on the service range weight matrix M. After completing the local model training, they upload the test signal through the air calculation algorithm to obtain the power compensation coefficient. After preprocessing and encoding the local model, the time-frequency synchronization is synchronized and the local model is uploaded through the air calculation algorithm and the global model is aggregated. The main purpose of the air calculation is to enable the core server to calculate a class of so-called nominal functions by combining the data transmitted simultaneously from all mobile agents.

[0116] Step 2 specifically includes the following steps:

[0117] Step 2.1: All mobile agents are based on the loss function f n (w n ) Perform local model training;

[0118] Step 2.2: After all mobile agents complete local model training, they use the air computing algorithm to aggregate local models. The global model after the t-th round of aggregation is w t for:

[0119]

[0120] Among them, μ t is the learning rate, A t is the total number of mobile agents selected for transmission in round t, is the label chosen by the nth mobile agent in the tth round of communication, when σ is a given threshold; is the power compensation coefficient, is the denoising coefficient of the core server, is the channel gain of the nth mobile agent in the tth communication round, for The Hessian matrix of ; the coefficients need to take into account channel inversion and power alignment to satisfy the paradigm of air calculation. The optimal solution is:

[0121]

[0122] Step 2.3: The core server receives the signal of the global model, which is expressed as

[0123]

[0124] in, is the local model of the nth mobile agent after the tth round of training,

[0125] P n,t is the maximum transmission power of the nth mobile agent, is the post-processing function;

[0126] The core server expects to receive the following information:

[0127]

[0128] Among them, f n (w n ) is the loss function of the nth mobile agent;

[0129] Then the output of the integrated aerial federated learning system on the core server side is

[0130]

[0131] Where: s t Receives signals from the global model for the core server.

[0132] The details of the Air-FedRS algorithm for region selection are as follows:

[0133]

[0134]

[0135] Step 3: After obtaining the global model from step 2, the core server calculates the global loss, updates the service range weight matrix M based on the global loss and the locations of all mobile agents, and finally broadcasts the global model. Step 3 specifically includes the following steps:

[0136] Step 3.1: The core server calculates the global loss function

[0137] Step 3.2: Service range weight matrix M. After each round of aggregation, if the new global model loss F t (w) is less than or equal to the global model loss F of the previous round t-1(w), then give the updated position information (ρ n ,θ n ) corresponds to the elements in the service range weight matrix M multiplied by a reward coefficient α. After each round of aggregation, if the new global model loss F t (w) is greater than the global model loss F of the previous round t-1 (w), then give the updated position information (ρ n ,θ n ) corresponds to the element in M ​​multiplied by a penalty coefficient β, that is

[0138]

[0139] Among them, α<1,β>1.

[0140] The present invention requires estimating the position and transmission power of all mobile agents, and therefore uses an NNG-based device sensor. The present invention processes the superimposed signals in the air calculation by referring to the signal processing method of the MUSIC algorithm. The specific process is as follows:

[0141] (1) Receive signal x, x is expressed as:

[0142]

[0143] Where L is the number of signals. a(θ l ) is the steering vector of the plane wave. s l is the source signal. n is the zero-mean Gaussian noise vector. (2) Calculate the sampling covariance matrix R:

[0144]

[0145] Where V is the sampling window size. H is the conjugate transpose of the received signal. The MUSIC algorithm assumes the orthogonality of the signal and noise in the subspace. That is,

[0146]

[0147] Ideally, the steering vector in the signal subspace is orthogonal to the noise subspace; i.e.

[0148] a(θ l )·Q n =0

[0149] (3) Calculate the optimal angle θ * ,Depend on

[0150]

[0151] Inspired by traditional MUSIC, the present invention adopts the same method to extract the features of the superposition signal. Matlab is used for scene modeling. Given a rectangular antenna array of size (4,4). It serves as the receiving antenna array of the core server. It actively receives the superposition signal calculated from the air according to the time node. The core server then stores and calculates the sampling covariance matrix of the received superposition signal. This has two advantages: (1) Using the sampling covariance matrix does not require limiting the sampling window to a fixed configuration; (2) Calculating the correlation between channels frees AoA estimation from the prior knowledge of signal type or modulation method.

[0152] The total number of mobile agents that upload local models in each round is unknown. This makes the traditional regression task model unapplicable in the scenario proposed in this application. If a regression task method is adopted, a threshold must be set to determine the number of arrival angles. In the case of low SNR, the accuracy will drop sharply. And the complexity of the data set increases as the step size of the direction angle segmentation decreases. Therefore, the present invention uses an NNG user sensor to solve this problem.

[0153] The NNG user sensor consists of a user quantity classifier and N max The user number classifier is a classification task model. It analyzes the signal and outputs the number of users Nuser. Each number of users then corresponds to a user-aware regressor.

[0154]

[0155] Table 1. Agent number estimator

[0156]

[0157]

[0158] Table.2.Device Perception Estimator Illustration

[0159] This regressor can predict the pitch angle and the corresponding signal arrival power based on the feature matrix. The output is a matrix of size 3×Nuser. The complete NNG perceptron structure is as follows Figure 3 The network layer details of the device number estimator are shown in Table 1. The network layer details of the device awareness estimator are shown in Table 2. The output of each layer is as follows:

[0160] output=dropout(ReLU(W·input+b))

[0161] Where dropout(·) is a temporary dropout layer. ReLU(·) is the activation function. W, input, and b are the weight, layer input, and bias coefficient, respectively.

[0162] In order to verify the feasibility of the NNG-based user perception estimator proposed in this paper, we compared it with the traditional MUSIC algorithm. Figure 4 As shown. First, the difference between the MUSIC algorithm and the NNG method in terms of mean square error under conditions of different numbers of users is compared. The ASE of MUSIC is significantly higher than that of the user sensor based on NNG. And when the ASE angle is close to 0, the performance of the MUSIC algorithm is the worst. On the contrary, the effect of NNG is the best. This is because the angle is too small, and the characteristic differences of the signal transmitted at a small angle are relatively small. As a result, MUSIC is difficult to distinguish the optimal direction of the signal. NNG can learn the differences at these small angles and has a better effect. From this perspective, it can also be seen that NNG can also learn features under low signal-to-noise ratio conditions. At the same time, the present invention compares the running time of the two methods. Under the same simulation conditions, the execution time of the NNG method is one order of magnitude less than that of the MUSIC algorithm. This is because the MUSIC algorithm iterates all resolutions to find the closest angle. At the same time, in order to intuitively represent the prediction results of NNG, the present invention compares the predicted angle and the actual angle. It can be found that almost all angles are on the function. NNG has good estimation performance. Therefore, the device position estimator based on neural network has a better application prospect.

[0163] In order to demonstrate the effectiveness of the proposed ISCC Air FedRS Framework, the present invention assumes a core base station service area where a regional dataset exists. The comparison methods are: (1) NISCC Air-FL: an algorithm without power scheduling and user selection strategy. (2) Robust OTA FL: an Air-FL algorithm that considers CSI. To simplify the analysis, the present invention sets the partitioning only by the pitch angle. 18° is an area. Then, under one core server, there are a total of ten areas. Then, two areas are randomly selected as special areas where regional datasets exist. Let σ = 3, α = 0.8, β = 1.2. And the MNIST dataset is used as the dataset for model training. Figure 5(a) shows the relationship between the MSE and the number of MAs simultaneously uploading in a round. To control for variables, a power-free channel is assumed. The baseline is the MSE under conditions without transmission errors and regional datasets. ISCC is the proposed framework. MUSIC is a line graph of the MUSIC algorithm providing location information to FedRS. NISCC is a line graph of the MSE without the ISCC Air FedRS Framework. ISCC is the algorithm closest to the baseline. This is because NNG provides more accurate perception information to FedRS. This effectively reduces the gradient bias caused by regional datasets. The gap becomes more significant as the number of MAs increases. This is because the probability of all mobile agents originating from a specific region increases. ISCC also outperforms the RobustOTAFL algorithm because it also considers the influence of regional datasets. Figure 5 (b) shows how accuracy varies with SNR. A fading channel is introduced here. The difference between the ISCC Air FedRS Framework and the baseline stems from the discrepancy between estimated and actual power. Furthermore, when the SNR is low, noise significantly impacts the global model. NISCC, however, exhibits significant model distortion due to its lack of power compensation. ISCC is similar to the Robust OTAFL algorithm.

[0164] Finally, the present invention verifies the influence of adjusting the hyperparameters σ, α, and β on the ISCC Air FedRS Framework. Figure 6 shown. Figure 6 (a) and (b) show the impact of different values ​​of σ as the number of communication rounds increases. It can be seen that a too small σ leads to a slow decrease in the MSE. This is because errors can misjudge special areas and silence normal areas. A too large σ leads to excessive MSE fluctuations. This is because the ISCC Air FedRS Framework cannot quickly reduce the impact of special areas and silence them. Figure 6 (c)(d) shows the effect of different α, β on MSE and accuracy. When , the reward weight is greater than the penalty weight, the ISCC Air FedRS Framework learns faster but causes greater volatility. When the penalty weight is greater than the reward coefficient, the ISCC Air FedRS Framework learns more slowly but is more stable. This is because a heavier emphasis on reward increases the number of local models participating in model aggregation. However, this can make it easier to aggregate local models from unusual regions. A heavier emphasis on penalty weight reduces the influence of local models from unusual regions more quickly, but it can also more quickly classify normal regions as unusual. Therefore, tuning the hyperparameters σ, α, and β is crucial for the ISCC Air FedRS Framework.

[0165] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. An airborne federated learning system integrating synergy and computing, characterized by: The integrated aerial federated learning system includes: Model learning module: It consists of the Air-Federated Learning Algorithm for Region Selection (Air-FedRS) module; Device perception module: The device perception module is an NNG user sensor, which processes the superimposed signal to perceive the location and signal status of the mobile agent that uploaded the local model; Mobile Agent (MA): collects data, trains local models based on regional representative weights, and uploads local models to the core server after training to aggregate the global model; Core Server (CS): Detects the location and signal status of mobile agents, updates regional representative weights, and sends corresponding power compensation coefficients to all N mobile agents. After receiving all local models, the core server aggregates and broadcasts the global model, inputs the global model into the model learning module, and uses the regional selection air federated learning algorithm (Air-FedRS) for the next round of federated learning. The air federated learning algorithm for region selection (Air-FedRS) specifically includes the following steps: Step 1: Initialize the service range weight matrix M according to the service range and specific requirements of the core server, initialize the global model and broadcast it to all mobile agents participating in the aerial federated learning, and determine the total loss function; In step 2, all mobile agents in step 1 receive the global model and perform local model training based on the service range weight matrix M. After completing the local model training, they upload the test signal through the over-the-air calculation algorithm to obtain the power compensation coefficient. After preprocessing and encoding the local model, they synchronize the time and frequency by uploading the local model through the over-the-air calculation algorithm and aggregating the global model. Step 3: After obtaining the global model in step 2, the core server calculates the global loss, updates the service range weight matrix M based on the global loss and the locations of all mobile agents, and finally broadcasts the global model.

2. The airborne federated learning system integrating synergy and computing according to claim 1, characterized in that: The NNG user sensor includes a classifier and multiple regression heads. The classifier is an agent number sensor, and the regression head is a single-agent sensor, a dual-agent sensor, or a multi-agent sensor. The agent number sensor is used to estimate the number of all mobile agents in the superimposed signal, and the single-agent sensor, the dual-agent sensor, and the multi-agent sensor are used to perceive the location and channel state information of all different mobile agents.

3. The in-flight federated learning system integrating synergy and computing according to claim 1 is characterized by: The NNG user sensor uses the representative features of the superimposed signals in the air federated learning to perceive the locations and signal states of different mobile agents, supplementing the perception function of the traditional air federated learning framework. The NNG user sensor processes the superimposed signals to perceive the location and signal state of the local model mobile agent, specifically including the following steps: Receive signal x Calculate the sampling covariance matrix R: Among them, V is the sampling window size, x H is the conjugate transpose of the received signal, L is the number of uploaded signals, l is the uploaded signal index, a is the channel gain, is the direction angle, n is the noise, s l is the source signal; The sampling covariance matrix R is input into the NNG user sensor for perception. The sampling covariance matrix R first enters the agent number sensor to obtain the number of all mobile agents N. t Then input the sampling covariance matrix R into the regression head to obtain the positions and signal states of all different mobile agents. If N t =1, then use a single agent sensor; N t =2, then use the dual agent sensor. If N t =3, then use three-agent perceptron, if N t =X, then use X proxy perceptron.

4. The airborne federated learning system integrating synergy and computing according to claim 1 is characterized by: The step 1 specifically includes the following steps: Step 1.1: Determine all mobile agents within the service range of the core server and determine the area division based on the pitch angle; Step 1.2: Initialize the service range weight matrix M: Define the regional representative weight of the nth mobile agent updated in the tth round as Where M(ρ n ,θ n ) is the core server corresponding to the location information (ρ n ,θ n ), where M is the service range weight matrix, and The range is (1, +∞). Before starting the first round of aggregation, all elements of the service range weight matrix M are initialized to 1; Step 1.

3. Determine the global loss function: global loss function for global model training Defined as Among them, N is the total number of mobile agents, N t is the number of mobile agents that upload local models in the tth round of aggregation, N t <N, is the set of communication rounds, T is the total number of communications, is the loss function of the nth mobile agent: in For a given global model parameter Under the premise that the nth mobile agent has n,i ,y n,i ), is the training set of each n-th mobile agent, x n is the input data accessible to the nth mobile agent, y n For the corresponding label is the collection of all inputs, A collection of all tags.

5. The airborne federated learning system integrating synergy and computing according to claim 4 is characterized by: Step 2 specifically includes the following steps: Step 2.1: All mobile agents are based on the loss function Perform local model training; Step 2.2: After all mobile agents complete local model training, they use the air computing algorithm to aggregate local models. The global model after the t-th round of aggregation is for: Among them, μ t is the learning rate, A t is the total number of mobile agents selected for transmission in round t, is the label chosen by the nth mobile agent in the tth round of communication, when σ is a given threshold; is the power compensation coefficient, is the denoising coefficient of the core server, is the channel gain of the nth mobile agent in the tth communication round, for The Hessian matrix of Step 2.3: The core server receives the signal of the global model, which is expressed as in, is the local model of the nth mobile agent after the tth round of training, P n,t is the maximum transmission power of the nth mobile agent, is the post-processing function; The core server expects to receive the following information: in, is the loss function of the nth mobile agent; Then the output of the integrated aerial federated learning system on the core server side is Where: s t Receives signals from the global model for the core server.

6. The airborne federated learning system integrating synergy and computing according to claim 5, characterized in that: The step 3 specifically includes the following steps: Step 3.1: The core server calculates the global loss function Step 3.2, the service range weight matrix M, after each round of aggregation, if the new global model loss Less than or equal to the global model loss of the previous round Then give the updated position information (ρ n ,θ n ) corresponds to the elements in the service range weight matrix M multiplied by a reward coefficient α. After each round of aggregation, if the new global model loss Greater than the global model loss of the previous round Then give the updated position information (ρ n ,θ n ) corresponds to the element in M ​​multiplied by a penalty coefficient β, that is Among them, α<1,β>1.

7. The airborne federated learning system integrating synergy and computing according to any one of claims 1 to 6, characterized in that: The airborne federated learning method of the integrated airborne federated learning system for synergy and computing specifically includes the following steps: Step (1): The core server initializes the global model parameters The service range weight matrix M, and then the global model parameters And the service range weight matrix M is broadcast to all mobile agents; Step (2): All mobile agents receive the global model parameters Then select the local model to train and get the local model Step (3): All mobile agents upload test signals to the core server, sense the user's channel status information through the NNG user sensor, and then calculate the power compensation coefficient. Step (4): Preprocess and encode the local model and synchronize the time and frequency to upload the local model and aggregate the global model through the air computing algorithm. Step (5): The core server receives the superimposed signal, decodes and post-processes it to obtain the global model. At the same time, the positions of all mobile agents are obtained by calculating the sampling covariance matrix R of the superimposed signal and using the NNG user sensor; Step (6): Calculate the global loss function through the global model And update the service range weight matrix M according to the locations of all mobile agents that uploaded the local model; Step (7): The core server broadcasts the new global model

Images