Air federated learning method for 6G wireless network communication perception integrated heterogeneous data

Abstract

The invention relates to an air federated learning method for 6G wireless network communication perception integrated heterogeneous data, which comprises the following steps of: initializing a scene, receiving an antenna signal based on an antenna parameter server of an antenna worker, and carrying out orthogonal frequency division multiplexing and non-orthogonal multiple access demodulation iteration to update global model parameters; the updating parameters are broadcasted to all antenna workers through a shared link and are used for executing local iteration updating, and then accumulation model updating is transmitted to an antenna worker antenna parameter server; setting an objective function which minimizes an end-to-end loss function, and checking federated learning convergence on a time-varying channel in combination with a local update constraint to complete model update; in the global model iteration updating process, cyclic prefix OFDM transmission words are constructed, a parameter server (PS) antenna is deployed after an antenna worker cyclic prefix is removed from a receiving antenna, all received signals corresponding to character indexes of orthogonal frequency division multiplexing and non-orthogonal multiple access are aggregated, and a local updating average estimation value is obtained.

Description

Technical Field

[0001] This invention relates to the field of aerial federated learning technology, and in particular to an aerial federated learning method for heterogeneous data integrating 6G wireless network communication and sensing. Background Technology

[0002] The sixth-generation mobile communication system (6G) aims to build a new world of intelligent interconnection and digital twins, placing higher demands on network communication speed, sensing capabilities, connection density, intelligence level, and anti-interference performance. Against this backdrop, Integrated Communication and Sensing Technology (ISAC) enables wireless networks to act as high-precision, real-time sensors for environmental perception while transmitting information, achieving "communication as sensing." This transforms wireless networks from simple "information pipes" into "information and sensing converged infrastructure," laying the foundation for various innovative applications.

[0003] Network application data is heterogeneous, including communication data such as images and text from user equipment, as well as environmental perception data such as target distance and speed obtained from Channel State Information (CSI). These two types of data differ significantly in modality, dimensionality, and statistical characteristics. As a dual carrier of communication quality and environmental characteristics, CSI plays a crucial role. How to collaboratively utilize heterogeneous data, especially mining inter-carrier interference (ICI) information from CSI, to overcome the constraints of heterogeneity on model training, has become a challenge for 6G networks. Federated learning (FL), as a distributed machine learning approach, provides a feasible path by training the model locally on the terminal and uploading the model gradient to the server for aggregation, thus alleviating transmission bottlenecks while protecting data privacy.

[0004] In 6G scenarios, the traditional "worker-parameter server" architecture of federated learning faces severe challenges: frequent model updates consume significant wireless resources, leading to transmission bottlenecks; the non-independent identically distributed (Non-IID) nature of heterogeneous data easily causes slow model convergence or even divergence; and the relative motion between workers and parameter servers (PS) and dynamic environmental changes involve time-varying channel characteristics, causing Doppler shift and spread, disrupting orthogonality, and generating inter-carrier interference (ICI), thus impairing system performance. To address this, this invention employs Over-the-Air Federated Learning (OTA-FL), which integrates communication and computation, allowing multiple devices to transmit in parallel on time-frequency resources, enabling model updates to be naturally superimposed in the wireless channel and directly received by the server. This method offers dual advantages: on the one hand, it utilizes over-the-air computation (OTA) to improve spectral efficiency and aggregation speed; on the other hand, it provides efficient and unified aggregation for integrated sensing heterogeneous data, effectively suppressing interference and improving model convergence performance. Summary of the Invention

[0005] To address the problems existing in the prior art, the purpose of this invention is to provide an over-the-air federated learning method for heterogeneous data integrating communication and perception in 6G wireless networks. This method constructs a self-enhancing "perception-communication-learning" intelligent closed loop, suitable for multi-level scenarios from smart cities (macro) and smart factories (meso) to smart homes and human body sensing (micro), thereby unleashing the potential of 6G networks in practical deployments.

[0006] To achieve the above objectives, the present invention provides the following solution: An over-the-air federated learning method for heterogeneous data integrating communication and sensing in 6G wireless networks includes: Scene initialization, i.e., building an airborne federated learning system consisting of a single-antenna worker and a multi-antenna parameter server; Based on the antenna signal received by the antenna parameter server, the global model parameters are iteratively updated by orthogonal frequency division multiplexing and non-orthogonal multiple access demodulation, and the updated parameters are broadcast to all antenna workers through a shared link. After performing local iterative updates, the cumulative model updates are transmitted to the antenna parameter server. Set the objective function to minimize the end-to-end loss function, and combine local update constraints to verify the convergence of federated learning on the time-varying channel, and complete the model update; During the global model iterative update process, a cyclic prefix orthogonal frequency division multiplexing (OFDM) transmission word is constructed to counteract the effects of multipath channels. After the cyclic prefix is ​​removed from the receiving antenna, a parameter server (PS) antenna is deployed. The received signals corresponding to the character indices of all OFDM and non-orthogonal multiple access (NOMA) signals are aggregated to obtain the local update average estimate, which is used to update the global model parameters.

[0007] Optionally, receiving the antenna signal for OFDM and NOMA demodulation includes: Based on orthogonal frequency division multiplexing (OFDM), non-orthogonal multiple access (NOA) is introduced to form an OFDM and NOA air interface. This interface is used to divide the local update vector into multiple parts, each of which occupies multiple resource units as an independent multiple access word for transmission, generating a local update segment. This allows resource allocation to be shifted to intelligent adaptive resource entities in the multi-dimensional domain for communication-sensing integrated communication (ISAC). ; ; ; in, This represents the global iteration count. Number the workers in the system. To update part of the index locally, For the first Round global model parameters, This is a partial update segment. and Corresponding to the real part and the imaginary part, The imaginary unit, This refers to the number of ISAC resource units. This is the last real part of the vector component. For the last imaginary vector component, This is a transpose operation.

[0008] Optionally, constructing the cyclic prefix OFDM transport word includes: Set the cyclic prefix length, and combine the cyclic prefix length with the local gradient information of the antenna operator to construct the cyclic prefix OFDM transmission word: ; in, This represents the global iteration count. Number the workers. To update part of the index locally, For cyclic prefix OFDM transmission words, This refers to the number of ISAC resource units. The length of the cyclic prefix (CP) This is the last component of the transmitted word. and This is the intermediate component of the transmitted word.

[0009] Optionally, deploying the PS antenna includes: Perform on the elements of the received vector The discrete Fourier transform (DFT) of a point yields the frequency domain signal: ; in, This represents the global iteration count. To update part of the index locally, For PS receiving antenna index, For transforming the index, For element index, For frequency domain signals, This refers to the number of ISAC resource units. For global iteration The first received vector in the middle One component; Substitute the frequency domain signal into the received signal: ; in, This represents the global iteration count. To update part of the index locally, For PS receiving antenna index, For transforming the index, Number the workers. For channel monitoring index, The received signal after substitution. Additive noise term DFT, Number of channels monitored. For the corresponding link number Channel tap delay For a moment From workers To PS root antenna number Gain of secondary channel monitoring These are global parameter components; The received signal includes: the required signal component, channel noise, interference from other workers on the same resource unit, and inter-carrier interference; Define the interference gradient coefficient and inter-carrier interference term group coefficient for other workers on the same resource unit: ， ; ; in, This represents the global iteration count. and Index for different workers experiencing interference , For the receiving antenna index, This represents the total number of receiving antennas. For transforming the index, For the interference gradient coefficients of other workers on the same resource unit, For inter-carrier interference term group coefficients, and For a zero-mean random variable, and For its conjugate, For frequency response components; The PS antenna is deployed based on the fact that the interference gradient coefficients of the other workers on the same resource unit and the inter-carrier interference term group coefficients satisfy the corresponding upper bound of variance: ; ; in, This represents the global iteration count. Index for workers, To update part of the index locally, For transforming the index, Describe the trend of how the algorithm changes as the size of the input data increases. The variance of the interference gradient coefficients of other workers on the same resource unit. This is the upper bound of the variance of the inter-carrier interference term coefficients. Number of channels monitored. For signal variance It is a measure of the change over time.

[0010] Optionally, aggregating the received signals corresponding to all character indices of orthogonal frequency division multiplexing (OFDM) and non-orthogonal multiple access (NMO) includes: The received frequency domain signals corresponding to the character indices of all orthogonal frequency division multiplexing (OFDM) and non-orthogonal multiple access (NMO) signals are aggregated; the components of the received frequency domain signals are represented as follows: ; ; ; ; ; in, This represents the global iteration count. For the first received signal One portion, , For the desired signal, For equivalent noise, To avoid interference from multiple workers, Inter-carrier interference, , For channel frequency response, It is the conjugate transpose. , , , , To update the vector components, Additive noise term The DFT.

[0011] Optionally, obtaining the local update average estimate includes: Based on the components of the received frequency domain signal, the locally updated average estimate is obtained: ; in, This represents the global iteration count. For local iterative indexing, For transforming the index, To update the global average estimate, This is a local update of the average estimate.

[0012] Optionally, setting the objective function includes: ; in, For the first The empirical loss function for the local dataset of each worker. For the number of workers, For each worker The size of the local dataset possessed This represents the total amount of data in the system.

[0013] Optionally, setting the local update constraint includes: ; in, This represents the global iteration count. For resource unit index, For mathematical expectation constraints, For the first Round global model parameters, Initialize it. The optimal local update strategy for workers. For and Different worker indexes, These are the two components of the local update constraint.

[0014] Optionally, setting the local update constraint further includes: ; ; in, This represents the global iteration count. To account for cross-worker data bias, This represents the total number of antennas. The total number of workers, For time delay variables, For resource unit index, For the corresponding learning rate, For local iteration updates of the total number, Let be the dimension of the real vector space. For mathematical expectation symbol, The sign for variance is . for The upper realm, It is a measure of the change over time.

[0015] The beneficial effects of this invention are as follows: This invention introduces heterogeneous data integrating communication and sensing into federated learning, and realizes the transformation from "connection" and "sensing" to "environment cognition and understanding" through channel state information.

[0016] This invention solves communication bottlenecks through over-the-air computing, achieving a balance between communication efficiency and intrinsic privacy. It employs multiple receiving antennas, supports synchronous transmission between nodes, and completes model aggregation in the air, reducing aggregation latency and interference.

[0017] This invention uses global intelligent updates to guide node resource scheduling, forming a self-reinforcing closed loop. This transforms the network from a passive transmission channel into an intelligent agent that actively perceives the environment, engages in distributed collaborative learning, and optimizes in real time. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of an over-the-air federated learning method for integrated heterogeneous data in 6G wireless network communication and sensing, according to an embodiment of the present invention. Figure 2 This is a diagram of the wireless heterogeneous data federated learning structure according to an embodiment of the present invention; Figure 3 This is a wireless network over-the-air aggregation diagram according to an embodiment of the present invention; Figure 4 This is a hardware and software device diagram according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This embodiment discloses an airborne federated learning method for integrated heterogeneous data of 6G wireless network communication sensing, including: (1) scenario initialization, constructing an airborne federated learning system composed of a single-antenna worker and a multi-antenna parameter server; (2) global model distribution, collecting local updates, end-to-end loss, mini-batch samples, gradient estimation, and building a local training dataset; (3) local gradient calculation, using the local heterogeneous dataset, calculating model updates (gradients) through stochastic gradient descent; (4) model uploading and airborne aggregation, equipping multi-parameter server antennas to reduce the impact of interference terms; (5) global model update, signal reception and superposition, and extraction of aggregation results; (6) looping and optimization, considering learning rate, heterogeneous data and resource constraints for the end-to-end loss target; (7) convergence judgment, verifying the convergence of federated learning on time-varying channels, and improving privacy protection and generalization ability.

[0022] In step (1), a hybrid OFDM and NOMA access method is adopted. Each worker superimposes their local gradient onto the OFDM time-frequency resource grid and transmits it to the parameter server (PS) through a time-varying multipath multiple access channel. The time-varying channel (path loss, multipath delay, Doppler shift) is no longer a transmission obstacle, but a factor that can be sensed and utilized. The PS uses a multi-antenna array to combat channel fading. Channel time-varying characteristics destroy subcarrier orthogonality, and inter-carrier interference (ICI) affects transmission reliability.

[0023] In 6G system integration sensing and communication (ISAC), the communication network transforms into a distributed sensor network, enabling collaborative communication and sensing. The time-varying characteristics of the channel become a crucial source of sensing information; the distance, velocity, and orientation of a target object are embedded in the time delay, Doppler shift, and angle changes of the channel impulse response. Federated learning (FL) provides this collaboration. Edge devices train models using locally acquired real-time data (reflecting the latest channel state), and over-the-air (OTA) technologies aggregate model updates in the wireless domain, constructing a global model adapted to the time-varying environment and achieving a closed loop of communication-sensing-learning.

[0024] The PS acts as the coordination center to maintain the global model, while workers train using local data, collaborating iteratively through a "train-upload-aggregate-distribute" process. A blind transmitter design (no Channel State Information (CSI) at the transmitter, but CSI estimation at the receiver) reduces overhead. ISAC-aware information (such as user location and mobility status) optimizes federated learning resource allocation, such as dynamically excluding users in blind spots or adjusting beamforming directions. The non-independent and identically distributed (Non-IID) nature of the data leads to inconsistent local update directions, introducing "client drift," requiring fault tolerance and scheduling design to ensure training efficiency.

[0025] In step (2), the resource unit (such as OFDM symbols and subcarriers) has a dual function: transmitting data and sensing the environment as a probe signal. The parameter server (PS) receives signals through multiple antennas, performs joint OFDM and NOMA demodulation to update the global model, and broadcasts the parameters to all workers through an error-free link to start the next round of training. The system's end-to-end global loss function is a weighted sum of the local losses of each worker. In each round of global iteration, after receiving the global parameters, the worker performs multiple local iterations, calculates the stochastic gradient using mini-batch samples, and updates the local model according to the learning rate. At the same time, the learning rate is adjusted according to the batch size to balance the training dynamics.

[0026] After the global iteration ends, the worker sends the accumulated model update data to the PS, and updates the global model by calculating the average of local updates. It supports massive device access, adopts an OFDM+NOMA air interface scheme, and extends resource allocation from a single subcarrier dimension to a multi-dimensional domain. The local update vector is divided into multiple segments, each occupying a fixed number of ISAC resource units for transmission. Update information is carried by the real and imaginary parts, and zero-padding is used to meet transmission format requirements.

[0027] In step (3), each worker uses stochastic gradient descent (SGD) to calculate model updates based on the local heterogeneous dataset. Discrete Fourier transform and its inverse transform (DFT / IDFT) serve as tools for implementing OFDM waveforms, working in conjunction with non-orthogonal multiple access (NOMA): NOMA is responsible for multi-user resource allocation strategies, while DFT / IDFT is responsible for signal generation and resource mapping. In each global iteration, the worker transforms the local update vector into a time-domain signal and adds a cyclic prefix (CP) to construct the OFDM transmission word to counteract the effects of multipath channels.

[0028] All workers can transmit updates without constraints, regardless of power limitations and channel attenuation. Channel state information is introduced, further expanding into user scheduling and dynamic resource allocation mechanisms to address practical communication constraints. In OFDM systems, the cyclic prefix (CP) transforms the linear convolutional channel into a cyclic convolution to simplify equalization, maintain subcarrier orthogonality, and avoid interference when the delay spread does not exceed the CP length.

[0029] The channel employs a first-order autoregressive process to reflect the intensity of time-varying characteristics. At the multi-antenna receiver, the signal contains gradient information superimposed from multiple users, along with additive noise. The PS estimates the local update average by processing all received signals, which is used to update the global model, and the next iteration is initiated via broadcast. The OFDM time-frequency resource grid serves as the basic transmission carrier, while NOMA serves as an enhancement mechanism for implementing multi-user multiplexing on top of it.

[0030] In step (4), after removing the cyclic prefix at the receiving end, the received signal is subjected to DFT transformation to recover the frequency domain information. The frequency domain received signal can be decomposed into four parts: the desired signal component, channel noise, co-channel interference from other operators, and inter-carrier interference. Inter-carrier interference is mainly caused by the time-varying characteristics of the channel destroying the orthogonality of subcarriers, while multi-user superimposed transmission leads to mutual interference between operators.

[0031] Increasing the size of the PS antenna array suppresses interference during the aerial computation aggregation process, thereby ensuring the accuracy and convergence stability of the global model update in federated learning. While introducing interference, time-varying channels also provide a source of sensing information for integrated communication and sensing. An effective balance is achieved between utilizing channel variations for environmental perception and suppressing their impact on communication quality. Optimizing the receiving antenna configuration simultaneously improves the aggregation efficiency and communication reliability of federated learning without relying on additional spectrum resources.

[0032] In step (5), under time-varying channel conditions, inter-carrier interference (ICI) affects the accuracy of airborne computation aggregation. The mobility and high-frequency band usage of 6G networks exacerbate interference. The ISAC architecture transforms interference into a potential information source, and analyzing interference patterns can invert the target's motion state. To control system interference, this invention limits interference to a finite range of adjacent resource units and quantifies the time-varying intensity of the channel through autoregressive parameters. Adding PS antennas effectively suppresses interference, utilizing the local characteristics of interference to limit its spread. In the global model update stage, PS aggregates the superimposed signals received by multiple antennas, extracts the real and imaginary parts, performs normalization, and restores the local update average. Joint processing and compensation of each component enables gradient aggregation under time-varying channel conditions, providing a guarantee for reliable convergence of federated learning in the 6G dynamic environment.

[0033] In step (6), the end-to-end loss is minimized, a joint optimization is constructed, and the optimal local update strategy is solved under the triple constraints of communication, perception, and learning. The sum of weighted local losses is minimized, and the optimal solution reflects the data distribution characteristics: cross-device data deviation can quantify the degree of Non-IID, and IID data deviation approaches zero. The loss function is assumed to have both smoothness and strong convexity, controlled by parameters that determine the curvature of the function and the rate of gradient change, and parameters that reflect the strength of the convexity of the objective function, respectively. Both affect the convergence speed and stability. Devices adopt a uniform learning rate, and the model is updated iteratively through local stochastic gradient descent. Solution methods include decomposition-based alternating optimization, intelligent scheduling of deep reinforcement learning, distributed pairing of matching theory, and personalized expression of group learning.

[0034] In step (7), the upper bound of convergence consists of the initial state distance and the residual distance, and is affected by the learning rate, data distribution, and channel conditions. Non-IID data bias and inter-carrier interference (ICI) increase the residual term, delaying convergence and reducing the optimality of the solution. Channel time-varying factors affect the upper bound of convergence through interference, and their intensity is positively correlated with the autoregressive parameters. Increasing the number of iterations and decreasing the learning rate can alleviate some of the scaling term's influence, but the non-scaling term caused by channel noise and time-varying response still needs to be suppressed through antenna configuration and resource optimization. In terms of generalization ability, the noise of stochastic gradient descent helps the model converge to a flat minimum, improving prediction performance on unknown data. The integrated node device integrates communication, sensing, and computing functions, providing facility support for the deployment of federated learning in a 6G dynamic environment.

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] This embodiment discloses an aerial federated learning method for heterogeneous data integrating 6G wireless network communication and sensing, which can be deployed at network nodes such as base stations, access points, drones, and smart roadside units. Figure 1 As shown, it includes the following steps: (1) Scenario initialization: Construct an airborne federated learning system consisting of a single-antenna worker and a multi-antenna parameter server. In the federated learning system, each worker (computing node) calculates the model gradient using its local dataset. Using an access method (such as a combination of OFDM and NOMA), the gradient signal is superimposed on the OFDM time-frequency resource grid and transmitted to the parameter server (PS) via a time-varying multipath fading multiple access channel. In the 6G environment, time-varying channels are not only a transmission obstacle but are also used for sensing. Path loss, multipath delay, and Doppler shift change rapidly over time due to user equipment movement, scatterer motion, and the use of high-frequency carriers (such as terahertz). This change requires real-time adaptation of communication and sensing strategies. This invention assumes that the worker does not possess channel state information, and the PS is configured with a multi-antenna array to combat channel fading. The time-varying nature of the wireless channel disrupts the orthogonality of OFDM subcarriers, and inter-carrier interference (ICI) poses a challenge to system transmission.

[0037] This invention achieves synergy between communication and sensing functions by sharing wireless signals and hardware facilities. The time-varying characteristics of the channel are not only a factor causing transmission interference but also a source of sensing information: parameters such as the distance, radial velocity, and azimuth of the target object are all contained in the time delay, Doppler shift, and angle changes of the channel impulse response. Federated learning (FL) provides efficient collaboration, with edge devices collecting real-time data (reflecting the latest channel state) locally to train their local models. Through over-the-air technology, model updates from each device are directly aggregated to the PS in the wireless domain, forming a globally intelligent model that rapidly adapts to the time-varying environment, thus realizing a communication-sensing-learning closed loop.

[0038] Consider a A wireless federated learning (FL) system consists of one user and one base station (BS) co-located with the server. The parameter server (PS) acts as the coordination and aggregation center, responsible for maintaining, updating, and distributing globally shared machine learning models. Each worker (compute node) acts as a local training node, training its own model using its private data. Through an iterative cycle of training-upload-aggregate-distribute, a global model is jointly trained without sharing the original data. Figure 2 This demonstrates a wireless over-the-air federated learning system (OTA-FL), in which workers are numbered as follows: Every worker Having a scale of The local dataset, the total system data volume is Each worker sends its calculated local gradients to the PS via a hybrid OFDM+NOMA access, and then receives and aggregates the gradient information from each node directly through over-the-air computation. The data held by different workers exhibits non-independent and identically distributed (Non-IID) characteristics, and the data patterns and amounts differ between devices, leading to inconsistent local model update directions and thus introducing "client drift" on the server side.

[0039] In wireless environments, acquiring complete Channel State Information (CSI) at the transmitter incurs overhead. A blind transmitter design is employed, eliminating reliance on CSI at the transmitter and instead using CSI estimation at the receiver. Integrated Communication and Sensing (ISAC) combines communication and environmental awareness, utilizing sensed information (such as user location, movement speed, and signal obstruction) to optimize resource allocation. For example, if a user is detected moving into a signal blind zone, they can be excluded from the current training round in advance. Furthermore, based on the spatial distribution of sensed users, beamforming directions can be dynamically adjusted to optimize over-the-air (OTA) aggregation coverage. CSI directly reflects the signal propagation environment, including the amplitude, delay, and angle of arrival / departure angle of multipath components, corresponding to the distance, motion state, and orientation information of objects in the environment.

[0040] (2) Global model distribution, collection of local updates, end-to-end loss, mini-batch samples, gradient estimation, and construction of a local training dataset; network resources are viewed as multifunctional fusion units. A resource unit (such as a set of OFDM symbols and subcarriers) is simultaneously or time-division multiplexed: it transmits data to users to realize communication functions, and also serves as a detection signal to realize the surrounding environment perception through reflected signals. Let... This indicates the receiving antenna index of the PS. This represents the total number of antennas. The PS (Power Controller) combines the received antenna signals, performs OFDM+NOMA demodulation, and updates the global model parameters. After each iteration, the PS broadcasts the updated parameters to all workers via a shared link, initiating the next round of training.

[0041] make It is the local model update vector of the worker. Indicates the first individual worker The empirical loss function for the local dataset. The total data volume of the entire system is... The amount of local data for each worker is The system's end-to-end global loss function is .set up This represents the global iteration count. For the first Round global model parameters. In each round of global iteration, PS transmits the parameters through an error-free link. Broadcast to workers. Each worker receives the global parameters and executes them. Local iteration update ( (For local iteration indexes). The gradient indicates the direction in which the loss function grows fastest. Let... For the first In the first round of global iteration Workers during sublocal iterations Stochastic gradient estimation, For the corresponding learning rate, Mini-batch samples are taken from the local data. The iterative rule for local model updates is as follows: The initial state of the local iteration is: Using small batches introduces significant gradient noise, which improves the model's generalization ability but also leads to instability in the convergence process. Using large batches makes training more stable and allows for parallel computation to accelerate the process, but weakens generalization performance and can lead to local optima. To maintain the stability of gradient noise, a strategy of adjusting the learning rate and batch size synchronously is adopted (when the batch size is doubled, the learning rate is also approximately doubled) to maintain dynamic balance in training.

[0042] Global Iteration At the end, each worker completed... After each local iteration update, the cumulative model update is transmitted to the PS, i.e. In an ideal, noise-free, and interference-free environment, the global model update rule of PS is as follows: (use (replace), among which This represents the average of the partial updates performed by all workers. .

[0043] To achieve efficient access for massive numbers of devices in 6G, Non-Orthogonal Multiple Access (NOMA) is introduced on top of Orthogonal Frequency Division Multiplexing (OFDM) to form an OFDM+NOMA air interface, supporting large-scale connection demands. To update part of the index locally, This represents the number of ISAC resource units. The local update vector will be used. Divided into Parts, each part occupies Each resource element is transmitted as an independent multiple access word. The partial update segment is... The real part and the imaginary part are defined as follows: ; ; in, Zero-padding is performed when necessary, ensuring the vector length is... The requirements. Let... For resource unit index, Representing vectors The One element, Resource allocation has shifted from the traditional single-dimensional subcarrier allocation in OFDM to intelligent adaptive resource entities oriented towards ISAC in a multi-dimensional domain.

[0044] (3) Local gradient calculation: Using local heterogeneous datasets, model updates (gradients) are calculated through stochastic gradient descent (SGD). Figure 3 Stochastic Gradient Descent (SGD) is employed, accepting noisy but unbiased gradient estimates in exchange for computational efficiency and generalization ability. In the Adaptive Momentum method, the current gradient and update direction are considered simultaneously during parameter updates by simulating physical momentum effects, suppressing zigzag oscillations. Discrete Fourier Transform (DFT) and its inverse transform (IDFT) are synergistic with Non-Orthogonal Multiple Access (NOMA) technology. NOMA is implemented in 6G networks, handling signal generation, resource mapping, and receiver design related to DFT / IDFT. In the... The next global iteration, based on the worker Local update vector Construct a multiple access signal. For complex fields, express A complex vector space. For the gradient vector... conduct Point IDFT, its first elements ( )for To counteract the effects of multipath channels, assume... This is the length of the cyclic prefix (CP). Combined with the worker... Based on local gradient information, construct a cyclic prefix OFDM transmission word: transmit cyclic prefix OFDM (CP-OFDM) word. All workers perform local model updates. The system has ample wireless resources, and power limitations and channel attenuation are not considered in each iteration.

[0045] 6G networks place high demands on the Cyclic Prefix (CP). As a guard interval added to the front end of OFDM symbols, the CP serves two purposes: first, it combats multipath delay spread by transforming linear convolutional channels into cyclic convolutional channels, simplifying frequency domain equalization; second, it maintains inter-carrier orthogonality, avoiding inter-carrier interference (ICI) when the delay spread is less than the CP length. In the next global iteration, let Update the segment index locally. For channel monitoring index, For time indexing, For time delay variables, Number of channels monitored. Definition. For a moment From workers To PS root antenna number Gain of secondary channel monitoring For the corresponding link number The channel tap delay and the channel impulse response are: ,in For a unit pulse. In the first... In this global iteration, the channel monitoring gain follows a first-order autoregressive model. ,in It has a mean of zero and a variance of . Independent complex random variables, The channel parameter reflects the correlation strength between consecutive samples.

[0046] In the One receiving antenna ( After removing the cycle prefix, update the segment index locally. Global iteration The first received vector in the middle elements ( )for ,in It is an additive noise term, following a cyclic symmetric complex Gaussian distribution, with its real and imaginary parts being independent and having equal power. Within the next global iteration, let For partial updates Average estimate, by processing all The received signal corresponding to each OFDM+NOMA character index is obtained. Then, the global model parameters are updated. Update parameters in the next global iteration This invention shares information with all workers via broadcast. It performs NOMA overlay on the time-frequency resource grid of OFDM, using DFT / IDFT as underlying tools for OFDM waveform generation and demodulation; NOMA is built on the physical layer to achieve multi-user resource reuse.

[0047] (4) Model uploading and over-the-air aggregation are performed using a multi-parameter server (PS) antenna to reduce the impact of interference terms; after removing the cyclic prefix (CP) at the receiver, the model is uploaded and aggregated in the air. The next global iteration updates the segmented index locally. By analyzing the first received vector element conduct Point DFT yields the corresponding frequency domain signal. For the transform index... After the receiver removes the CP, the first Within the next global iteration, adopt The point receiving vector of the first One entry The frequency domain signal is obtained by the DFT, and has Substituting the received signal, we can obtain... ,in It is an additive noise term. The DFT.

[0048] No. Within the next global iteration, let To update a portion of the index locally, the received signal is... The signal is decomposed into three parts: namely , specifically, ① , ② , ③ ,in , indicating from workers To the The channel frequency response of each PS antenna. Specifically, when... At that time, the channel remains unchanged. Let... For channel monitoring index, there are .

[0049] In the In the next global iteration, consider The frequency domain received signal is located on the root receiving antenna. Let... For character indexes, for transform indexes The result is obtained by combining complex conjugates. ,in A zero-mean random variable with variance ,here Further derivation yields... In the case of time invariance The variance is simplified to .

[0050] The movement of wireless devices causes channel time-varying, resulting in Doppler shift and spread, which disrupts the orthogonality between resource elements in a multi-carrier system, introduces inter-carrier interference (ICI), reduces communication reliability, and affects the accuracy of gradient aggregation. The elements of the gradient vector received at PS interfere with each other due to channel time-varying, thus affecting the quality of the global model update. In the next global iteration, let... For character indexing, from The received signal from a single receiving antenna is aggregated from four components: the desired signal component, channel noise, and other components (blind workers) within the same resource unit. Inter-carrier interference (ICI), i.e. The specific expression is as follows: ① ; ② ; ③ ; ④ ; According to the law of large numbers, when the number of trials is sufficiently large, the sample mean converges to the population mean. Therefore, the number of antennas at the receiving end... Under these conditions, the received signal converges to .

[0051] exist In the second global iteration, signals from different workers are combined at the receiver, resulting in the... Interference exists in the transmitted signal segment. Considering the worker... For workers Interference, defining the interference gradient in the received signal. The coefficient is , . The variance is Approximately Due to system existence There are several such interference terms, each with a mean of zero and a variance equal to the number of PS antennas. Inversely proportional to the number of receiving antennas, the interference term asymptotically approaches zero. Deploying a sufficient number of PS antennas effectively suppresses the aggregation effect of interference terms in the received signal, thereby avoiding destructive impacts on system performance.

[0052] (5) Global model update, signal reception and superposition, and extraction of aggregation results; Inter-carrier interference (ICI) is caused by frequency shift and phase noise disrupting the orthogonality between subcarriers. 6G network mobility and high spectrum usage exacerbate interference: In ultra-high frequency bands such as terahertz and millimeter waves, even at relatively low speeds, the absolute value of Doppler shift will still increase significantly, causing frequency spread and rapid channel time-varying; in high-speed rail transit and ultra-high-speed scenarios of air platforms, Doppler shift causes the channel to exhibit dual selective fading in both time and frequency dimensions; at the same time, the phase noise brought by high-frequency oscillators is also more prominent, directly exacerbating interference.

[0053] This invention limits the sources of inter-carrier interference (ICI) to a finite range of adjacent resource cells. For transform indexes... Assuming the first Each resource unit is affected by the combined forces on both sides. Interference from adjacent resource units. Since ICI power exhibits local concentration in the frequency domain, most interference originates from neighboring units; energy leakage from more distant resource units is negligible. The channel's time-varying speed is determined by parameters in the first-order autoregressive model. Characterization. To establish the interference range. With channel parameters Based on numerical correlation, an OFDM+NOMA character structure is designed: information is embedded only in a single resource unit, while the remaining units remain idle. This character is transmitted via a time-varying channel, and the corresponding interference range can be determined by measuring the energy leakage level on adjacent resource units. In wireless channels, smaller autoregressive parameters This corresponds to slower channel variations, thus limiting interference leakage between adjacent resource units, resulting in a smaller effective interference range. The weaker the time-varying nature of the channel, the more limited the spread range of ICI, and the lower the system interference.

[0054] In the ISAC architecture, high-precision velocity sensing can be achieved by accurately measuring the ICI component in the received signal, thus retrieving the radial velocity of the target. In the... Sub-global iteration, considering transmitted signal segments Consider each Group different inter-carrier interference (ICI) terms, and define each group of inter-carrier interference (ICI) terms. The corresponding coefficient is For a fixed gradient value, in the transmitted signal segment In the middle, the first The absolute value of the mean of the inter-carrier interference (ICI) term for each worker satisfies The upper bound of variance is satisfied. Therefore, as the ratio of interfering adjacent resource units to the total number of resource units decreases, the total system interference is limited to a finite range. Because... The variance includes with A proportional term, therefore a large number of receiving antennas are used (i.e., the number of antennas is increased). This can effectively suppress inter-carrier interference (ICI). In complex moving environments, the different Doppler components generated by multiple scatterers form ICI patterns, which help to perceive the dynamic characteristics of the environment. To convert interference into sensing information, it is necessary to have the ability to accurately separate the interference components from the mixed signal and establish a reliable mapping between them and the moving state.

[0055] Based on a real-time updated global model, the system can dynamically and intelligently adjust the entire network's communication strategy and sensing tasks: at the communication layer, it adaptively manages beams and allocates resources; at the sensing layer, it performs waveform selection and target tracking. Through closed-loop adjustment, it can continuously maintain high reliability of communication and high accuracy of sensing in time-varying channel environments. The next global iteration, through the transmission signal segments and transformation index The corresponding received signals are aggregated to obtain the locally updated average value. The estimator, specifically expressed as: , The estimation method extracts the received signal. The real and imaginary parts are used to perform... This involves normalizing the coefficients to restore the local update average value.

[0056] Over-the-air (OTA) aggregation of time-varying channels brings about a shift in communication resource allocation, moving from allocation per device to allocation per task. This prevents communication overhead from increasing linearly with the number of participating devices, thus improving spectral efficiency. This invention uses time-varying channels to reduce the impact of inter-carrier interference (ICI) on algorithm convergence, accelerating the training process. Assuming local updates of segment numbers... Number of resource units Therefore, superscripts are omitted thereafter. In global iteration In China, The signals received by the root antenna are aggregated. The first frequency domain signal received... Components (transformation index) ) can be re-represented as , The components are defined as follows: ① Expected signal ; ② Equivalent noise ; ③ Interference from multiple workers ; ④ Inter-carrier interference ; Based on the four signal components, in the... In the next global iteration, the required signal estimate is: .for ,have Over-the-air aggregation and global update involve three key steps: First, the base station receives the superimposed signal results from multiple nodes and channel noise; second, it extracts aggregation information from the superimposed signals; and finally, after appropriate scaling and correction, it updates the global model using the aggregation results.

[0057] (6) Iteration and optimization: For the end-to-end loss objective, consider the learning rate, heterogeneous data and resource constraints; This invention balances learning performance, communication awareness, interference handling, and resource constraints. It employs an end-to-end loss function to find the optimal local update strategy for workers. Minimize the objective function It is constrained by communication layer constraints, perception layer constraints, and learning layer constraints. Local model parameters are quantized before uplink transmission to minimize end-to-end loss. Minimum value. ,make For individual workers The optimal value of the loss function. To characterize the bias in the data and its heterogeneity across devices, cross-worker data bias is defined. ,in For non-IID data distributions, The value is higher for IID data with a sufficient number of samples. Approaching zero. For Non-IID data, aerial computation aggregation can be performed, but it can be combined with FL to form a synergy.

[0058] The adaptive learning rate adjusts the step size individually for each parameter. Based on historical gradient information, a smaller learning rate is used for parameters with larger gradients to ensure stability, while a larger learning rate is assigned to parameters with smaller gradients to accelerate convergence. In global iteration... All workers adopt a uniform learning rate. In the first At the individual worker level, model updates are achieved through local iterations. Execute and update the rules as follows , Assume the loss function for each worker. All Smooth and Strongly convex. Simultaneously, the expected square of the stochastic gradient... Norm is bounded, that is , Let represent the Euclid norm of the vector. The objective function also possesses smoothness (due to...). Control) and strong convexity (by (To ensure smoothness, a larger step size can be used, and strong convexity ensures an extremely fast linear convergence rate; if both cannot be satisfied at the same time, the algorithm performance will be limited.) Determines the curvature of the function: The loss function is relatively flat, and the gradient changes slowly; The larger the value, the steeper the function shape and the more dramatic the gradient changes. Controlling convexity strength: The size is relatively small, the objective function is close to a normal convex shape, and the overall shape is relatively flat; The function is relatively large, exhibiting a "deep bowl" shape, and possesses convergence guidance capabilities.

[0059] The solution to the aforementioned mixed-integer nonlinear programming problem is as follows: Iterative decomposition breaks down the original problem into sub-problems such as resource allocation, learning rate, and interference avoidance, gradually approaching the global optimum through alternating optimization; deep reinforcement learning constructs workers as intelligent agents, autonomously learning multi-dimensional resource joint scheduling by observing CSI, data distribution, and perceiving needs; matching theory establishes bidirectional preferences between workers and PSs, achieving dynamic pairing optimization through a distributed mechanism; and swarm learning allows heterogeneous nodes to maintain differentiation, balancing individualization and global consistency through soft aggregation methods such as knowledge distillation. These methods can be implemented independently or collaboratively.

[0060] (7) Convergence Judgment: Verify the convergence of federated learning on time-varying channels to improve privacy protection and generalization ability. Stochastic Gradient Descent (SGD) calculates the gradient of a single training sample at each parameter update step, adjusting the model parameters in the opposite direction of the gradient, achieving efficient updates through minimal samples. Mini-batch gradient descent balances stability and efficiency, providing a compromise for channel information processing. Facing the challenges of time-varying environments, it is necessary to consider performance limits under non-ideal conditions, develop novel compensation methods based on channel prediction and learning, and provide an upper bound on the convergence speed. Consider the number of global iterations. Learning rate Each global iteration contains This local update has the objective function as follows: Strongly convex. Assume the learning rate satisfies: . It is a partial update by the worker, according to the objective function. There are local update constraints: ,in , .here, yes The upper realm, Given autoregressive coefficients The number of resource units for inter-carrier interference (ICI). The computation considers the joint uncertainty arising from stochastic gradients and the randomness of the wireless channel. Learning rate scheduling includes a warm-up mechanism used in the early stages of training, gradient descent with a gradually decreasing learning rate according to predetermined rules, and annealing that follows a periodic adjustment based on a cosine function, ensuring that model training maintains both efficiency and stability.

[0061] The process of recovering the local update average from all workers at the PS end is subject to three types of interference: channel noise, blind worker interference, and inter-carrier interference (ICI). These interferences have upper bounds, reflected in the gradient variance expression. Interference must be considered when evaluating performance; otherwise, the actual system performance cannot be accurately reflected. Furthermore, increasing the number of workers participating in the learning process helps lower the performance upper bound and improve learning effectiveness and stability. After global iteration, based on... Smoothness can be used to derive the upper bound of convergence for the time-varying channel FL. This upper bound consists of the initial error and residual terms. While some residuals weaken with the learning rate decay, components caused by channel time-varying and noise are unaffected by learning rate scaling. Increasing the number of server antennas can effectively suppress such residual biases and improve convergence accuracy.

[0062] In terms of generalization ability, SGD exhibits a significant preference for flat minima. From the perspective of stochastic processes and statistical physics, SGD shows a tendency to migrate from steep regions to flat valleys in the early stages of training. The heavy-tailed distribution of gradient noise drives the model parameters to stabilize in a wide, flat region; this dynamic characteristic helps improve the model's predictive performance on unknown data. Combined with intelligent and adaptive optimizers, the learning strategy is dynamically adjusted by sensing the training state, significantly reducing the reliance on manual parameter tuning.

[0063] like Figure 4 As shown, the intelligent node is the core device of the system, integrating a hardware platform and software system. The hardware includes a multimodal data acquisition and processing unit, an over-the-air computing signal processing unit, and network interfaces and synchronization units, collectively supporting the coordination of communication, sensing, and computing functions. The software layer encompasses a sensory resource intelligent scheduler, a heterogeneous client engine, an over-the-air aggregation server engine, and a closed-loop controller, enabling dynamic resource allocation, local training, and over-the-air aggregation. This device overcomes the limitations of traditional methods that separate communication, sensing, and learning functions through hardware sharing, signal multiplexing, and process integration. Over-the-air computing combines model update transmission and aggregation computing into one, providing data transmission support for efficient, secure, and intelligent distributed learning in a 6G environment.

[0064] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An over-the-air federated learning method for heterogeneous data integrating 6G wireless network communication and sensing, characterized in that, include: Scene initialization, i.e., building an airborne federated learning system consisting of a single-antenna worker and a multi-antenna parameter server; Based on the antenna signal received by the antenna parameter server, the global model parameters are iteratively updated by orthogonal frequency division multiplexing and non-orthogonal multiple access demodulation, and the updated parameters are broadcast to all antenna workers through a shared link. After performing local iterative updates, the cumulative model updates are transmitted to the antenna parameter server. Set the objective function to minimize the end-to-end loss function, and combine local update constraints to verify the convergence of federated learning on the time-varying channel, and complete the model update; During the global model iterative update process, a cyclic prefix orthogonal frequency division multiplexing (OFDM) transmission word is constructed to counteract the effects of multipath channels. After the cyclic prefix is ​​removed from the receiving antenna, a parameter server (PS) antenna is deployed. The received signals corresponding to the character indices of all OFDM and non-orthogonal multiple access (NOMA) signals are aggregated to obtain the local update average estimate, which is used to update the global model parameters.

2. The over-the-air federated learning method for integrated heterogeneous data of 6G wireless network communication and sensing according to claim 1, characterized in that, Receiving the antenna signal for OFDM and NOMA demodulation includes: Based on orthogonal frequency division multiplexing (OFDM), non-orthogonal multiple access (NOA) is introduced to form an OFDM and NOA air interface. This interface is used to divide the local update vector into multiple parts, each of which occupies multiple resource units as an independent multiple access word for transmission, generating a local update segment. This allows resource allocation to be shifted to intelligent adaptive resource entities in the multi-dimensional domain for communication-sensing integrated communication (ISAC). ； ； ； in, This represents the global iteration count. Number the workers in the system. To update part of the index locally, For the first Round global model parameters, This is a partial update segment. and Corresponding to the real part and the imaginary part, The imaginary unit, This refers to the number of ISAC resource units. This is the last real part of the vector component. For the last imaginary vector component, This is a transpose operation.

3. The over-the-air federated learning method for integrated heterogeneous data of 6G wireless network communication and sensing according to claim 1, characterized in that, Constructing the cyclic prefix OFDM transport word includes: Set the cyclic prefix length, and combine the cyclic prefix length with the local gradient information of the antenna operator to construct the cyclic prefix OFDM transmission word: ； in, This represents the global iteration count. Number the workers. To update part of the index locally, For cyclic prefix OFDM transmission words, This refers to the number of ISAC resource units. The length of the cyclic prefix (CP) This is the last component of the transmitted word. and This is the intermediate component of the transmitted word.

4. The over-the-air federated learning method for integrated heterogeneous data of 6G wireless network communication and sensing according to claim 1, characterized in that, Deploying the PS antenna includes: Perform on the elements of the received vector The discrete Fourier transform (DFT) of a point yields the frequency domain signal: ； in, This represents the global iteration count. To update part of the index locally, For PS receiving antenna index, For transforming the index, For element index, For frequency domain signals, This refers to the number of ISAC resource units. For global iteration The first received vector in the middle One component; Substitute the frequency domain signal into the received signal: ； in, This represents the global iteration count. To update part of the index locally, For PS receiving antenna index, For transforming the index, Number the workers. For channel monitoring index, The received signal after substitution. Additive noise term DFT, Number of channels monitored. For the corresponding link number Channel tap delay For a moment From workers To PS root antenna number Gain of secondary channel monitoring These are global parameter components; The received signal includes: the required signal component, channel noise, interference from other workers on the same resource unit, and inter-carrier interference; Define the interference gradient coefficient and inter-carrier interference term group coefficient for other workers on the same resource unit: ， ； ； in, This represents the global iteration count. and Index for different workers experiencing interference , For the receiving antenna index, This represents the total number of receiving antennas. For transforming the index, For the interference gradient coefficients of other workers on the same resource unit, For inter-carrier interference term group coefficients, and For a zero-mean random variable, and For its conjugate, For frequency response components; The PS antenna is deployed based on the fact that the interference gradient coefficients of the other workers on the same resource unit and the inter-carrier interference term group coefficients satisfy the corresponding upper bound of variance: ； ； in, This represents the global iteration count. Index for workers, To update part of the index locally, For transforming the index, Describe the trend of how the algorithm changes as the size of the input data increases. The variance of the interference gradient coefficients of other workers on the same resource unit. This is the upper bound of the variance of the inter-carrier interference term coefficients. Number of channels monitored. For signal variance It is a measure of the change over time.

5. The over-the-air federated learning method for integrated heterogeneous data of 6G wireless network communication and sensing according to claim 1, characterized in that, Aggregation of received signals corresponding to character indices of all orthogonal frequency division multiplexing (OFDM) and non-orthogonal multiple access (NMO) signals includes: The received frequency domain signals corresponding to the character indices of all orthogonal frequency division multiplexing (OFDM) and non-orthogonal multiple access (NMO) signals are aggregated; the components of the received frequency domain signals are represented as follows: ; ； ； ； ； in, This represents the global iteration count. For the first received signal One portion, , For the desired signal, For equivalent noise, To avoid interference from multiple workers, Inter-carrier interference, , For channel frequency response, It is the conjugate transpose. , , , , To update the vector components, Additive noise term The DFT.

6. The over-the-air federated learning method for integrated heterogeneous data of 6G wireless network communication and sensing according to claim 5, characterized in that, Obtaining the local update average estimate includes: Based on the components of the received frequency domain signal, the locally updated average estimate is obtained: ； in, This represents the global iteration count. For local iterative indexing, For transforming the index, To update the global average estimate, This is a local update of the average estimate.

7. The over-the-air federated learning method for integrated heterogeneous data of 6G wireless network communication and sensing according to claim 1, characterized in that, Setting the objective function includes: ； in, For the first The empirical loss function for the local dataset of each worker. For the number of workers, For each worker The size of the local dataset possessed This represents the total amount of data in the system.

8. The over-the-air federated learning method for integrated heterogeneous data of 6G wireless network communication and sensing according to claim 1, characterized in that, Setting the local update constraint includes: ； in, This represents the global iteration count. For resource unit index, For mathematical expectation constraints, For the first Round global model parameters, Initialize it. The optimal local update strategy for workers. For and Different worker indexes, These are the two components of the local update constraint.

9. The over-the-air federated learning method for integrated heterogeneous data of 6G wireless network communication and sensing according to claim 8, characterized in that, Setting the local update constraint also includes: ； ； in, This represents the global iteration count. To account for cross-worker data bias, This represents the total number of antennas. The total number of workers, For time delay variables, For resource unit index, For the corresponding learning rate, For local iteration updates of the total number, Let be the dimension of the real vector space. For mathematical expectation symbol, The sign for variance is . for The upper realm, It is a measure of the change over time.

Images