Federal segmentation learning method and system based on cellular-free network

Abstract

The invention provides a federal segmentation learning method and system based on a cellular-free network, and the method comprises the following steps: building a cooperative service relation between a plurality of edge devices and a plurality of access points in the cellular-free network; when the neural network is trained, the neural network model is divided into a local model and a cloud model based on a segmentation layer, and the local model and the cloud model are respectively deployed on the plurality of edge devices and the cloud server; the edge device uploads intermediate data output by a segmentation layer in the local model to the cloud server through the collaborative service relationship; and the cloud server updates a global model based on the intermediate data and the cloud model, and issues the updated segmentation layer gradient to the edge device through the collaborative service relationship. According to the federal segmentation learning method and system based on the cellular-free network, on the basis of comprehensively considering segmentation layer selection, power distribution, bandwidth distribution and transmission quantization, tradeoff optimization of energy consumption and transmission delay is achieved.

Description

Technical Field

[0001] This invention belongs to the technical field of wireless communication, and in particular relates to a federated segmentation learning method and system based on non-cellular networks. Background Technology

[0002] In recent years, the rapid proliferation of mobile devices, wearable devices, and autonomous vehicles has generated massive amounts of data through wireless network systems, driving the rapid development of machine learning (ML) technology in related fields. However, traditional centralized machine learning frameworks have inherent drawbacks: user terminals need to upload the collected raw data completely to a cloud data center for unified processing and analysis. This model not only faces serious privacy protection issues but is also limited by network transmission bandwidth, leading to significantly increased communication latency and energy consumption. Especially in scenarios involving sensitive data, such as healthcare and financial payments, the centralized storage and transmission of raw data is more likely to trigger privacy breaches.

[0003] To address the aforementioned issues, Federated Learning (FL), as an emerging distributed machine learning framework, has garnered significant attention. Its core idea is to allow multiple edge devices to collaboratively train machine learning models without sharing raw data, thus protecting data privacy while improving model performance. Although FL offers advantages in privacy protection, its traditional framework still has limitations. Edge devices need to upload complete model parameters, which can lead to excessive computational and communication overhead in resource-constrained network environments, making high latency and high energy consumption unavoidable.

[0004] In terms of communication methods, cellular networks are gradually becoming an important direction for the future Internet of Things (IoT) due to their flexible deployment and low cost. Unlike traditional cellular networks, cellular networks eliminate cell division and achieve collaborative communication between user equipment (UE) and multiple access points (APs) by deploying a large number of distributed access points (APs) within a region. Their advantages include high reliability, uniform coverage, superior energy efficiency, and low latency. This distributed characteristic is highly compatible with the collaborative mode of light-emitting fiber (FL). Through coherent joint transmission, cellular networks can simultaneously provide services to multiple user equipment (UE), reducing latency, improving spectrum efficiency, and meeting the high-frequency communication needs of FL. At the same time, the cellular architecture based on massive MIMO (cell-free mMIMO) can provide stable, high-quality links, ensuring communication reliability during FL iteration.

[0005] Currently, various research efforts have been undertaken to explore FL (Flexible Learning) technology supported by non-cellular networks. For example, methods such as gradient compression and model pruning are used to reduce communication overhead; joint optimization of transmit power and data rate is employed to reduce uplink latency and energy consumption; and improved neural network backpropagation mechanisms and genetic algorithms are used for dynamic access point selection to enhance energy efficiency and system stability. While these methods have made some progress in optimizing traditional FL performance, they are still primarily designed around overall parameter transmission and have not adequately addressed the challenges posed by new distributed learning models.

[0006] Split-Federated Learning (SFL), a significant evolution of FL, offers a more resource-friendly solution. Its basic idea is to divide a deep neural network into a local model and a cloud model. The local model only computes the first few layers and uploads the intermediate feature data (smashed data) output from the splitting layers. The cloud server receives this data, performs the remaining computation and model aggregation, and returns the gradients from the splitting layers for the device to update the local model. This approach significantly reduces the computational burden on terminal devices, making it particularly suitable for resource-constrained scenarios such as smartphones and wearable devices. Furthermore, the amount of intermediate feature data is typically an order of magnitude smaller than the full model parameters, greatly reducing communication pressure. More importantly, the original data remains on the device, ensuring more thorough privacy protection. Therefore, SFL demonstrates unique advantages in scenarios sensitive to both privacy and latency, such as medical image processing and connected vehicle intelligence.

[0007] However, current research on federated segmentation learning (SFL) supported by non-cellular networks remains extremely limited. Existing methods mainly focus on parameter compression and link optimization in traditional fractional-process (FL), while the communication characteristics of SFL are fundamentally different. The transmitted object changes from complete parameters to intermediate feature data, whose data scale, distribution, and latency sensitivity are all different from those of parameters. Traditional gradient compression and model pruning strategies are difficult to directly transfer to SFL, resulting in a lack of effective optimization methods in non-cellular environments. Furthermore, the dynamic and heterogeneous nature of non-cellular networks further increases the complexity of the problem. The choice of segmentation layer directly determines the allocation of computational and communication overhead, and bandwidth and power allocation affect signal quality and energy consumption. While transmission quantization can reduce data volume, it may introduce model accuracy loss. These factors are coupled, and no systematic research has yet proposed a joint optimization strategy. Therefore, there is still considerable room for exploration in the systematic optimization of energy consumption control, communication efficiency, and latency balance in SFL within non-cellular network environments. For example, traditional federated learning optimization methods (such as gradient compression) cannot effectively handle the intermediate data transmission problem of the segmentation layer output in SFL, and do not consider the joint impact of the dynamics and heterogeneity of non-cellular networks on segmentation layer selection and bandwidth allocation.

[0008] This problem is particularly prominent in practical applications. In vehicle-to-everything (V2X) scenarios, vehicles rely on distributed learning for traffic flow prediction and autonomous driving assistance, but their computing and communication resources are limited, and high-speed movement leads to frequent changes in channel states. Without an effective segmentation layer selection and bandwidth allocation mechanism, intermediate data transmission will result in significant latency and energy consumption. In medical edge computing scenarios, mobile terminals need to process high-resolution images while protecting privacy. SFL can alleviate computational pressure, but if network resources are limited, transmitting intermediate features may also lead to excessive latency, affecting the real-time performance of diagnoses.

[0009] Although there has been some research on the integration of non-cellular networks and light-emitting fiber (FL), the optimization problem of light-emitting fiber (SFL) in non-cellular network environments has not yet been systematically solved. Existing optimization methods lack joint optimization strategies that address the characteristics of non-cellular networks and SFL communication modes.

[0010] How to achieve joint optimization of segmentation layer selection, power allocation, bandwidth allocation and transmission quantization in dynamic network environments, and achieve a balance between energy consumption and latency, is a key challenge that urgently needs to be overcome. Summary of the Invention

[0011] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a federated segmentation learning method and system based on non-cellular networks, which achieves a trade-off optimization between energy consumption and transmission latency by comprehensively considering segmentation layer selection, power allocation, bandwidth allocation and transmission quantization.

[0012] In a first aspect, the present invention provides a federated segmentation learning method based on a non-cellular network. The method includes the following steps: establishing a collaborative service relationship between multiple edge devices and multiple access points in a non-cellular network; for each edge device, obtaining the segmentation layer, bandwidth allocation, and transmission power of a neural network model; during neural network training, dividing the neural network model into a local model and a cloud model based on the segmentation layer, and deploying them respectively on the multiple edge devices and the cloud server; the edge devices uploading intermediate data output from the segmentation layer in the local model to the cloud server via the collaborative service relationship; the cloud server updating the global model based on the intermediate data and the cloud model, and distributing the updated segmentation layer gradient to the edge devices via the collaborative service relationship; the edge devices updating the local model according to the segmentation layer gradient, and dynamically adjusting the segmentation layer, bandwidth allocation, and transmission power to perform the next round of training of the neural network model until the neural network model converges.

[0013] In one implementation of the first aspect, establishing a collaborative service relationship between multiple edge devices and multiple access points includes the following steps:

[0014] For each edge device, an uplink is established with the multiple access points;

[0015] Obtain the large-scale fading coefficient of the uplink corresponding to the edge device and the access point;

[0016] For each edge device, a preset number of access points are selected from largest to smallest according to the large-scale fading coefficient to construct a service set;

[0017] When the access point belongs to the service set, it indicates that there is a service relationship with the corresponding edge device; when the access point does not belong to the service set, it indicates that there is no service relationship with the corresponding edge device.

[0018] In one implementation of the first aspect, the process of the edge device uploading intermediate data output from the segmentation layer in the local model to the cloud server via the collaborative service relationship includes the following steps:

[0019] The edge device uses delayed aggregation quantization to uniformly quantize the differences in the intermediate data and generates quantized data by backfilling the residuals.

[0020] The quantified data is uploaded to the cloud server via the collaborative service relationship.

[0021] In one implementation of the first aspect, the number of bits of the quantized data is d·b, where d represents the quantization dimension and b represents the quantization bit width.

[0022] In one implementation of the first aspect, the segmentation layer and the bandwidth allocation are adjusted based on the alternating direction multiplier method.

[0023] In one implementation of the first aspect, the transmission power is adjusted based on the generalized Dinkelbach method.

[0024] In one implementation of the first aspect, the neural network model converges when any of the following conditions are met:

[0025] 1) Reach the preset number of training rounds;

[0026] 2) The target value λτ+μ is less than the preset threshold, where τ represents the global transmission delay, and E k Let λ represent the energy consumption of the k-th edge device, and let λ and μ represent the tradeoff coefficients.

[0027] In one implementation of the first aspect, Where z k ,w k p represents the segmentation layer, bandwidth allocation, and transmission power allocation vectors of the k-th edge device, respectively. and Let z represent the segmentation layer of the k-th edge device. k Local computation and upload volume at that time, κ k , R k (p) and These represent the unit computational power consumption constant, RF transmit power constant, unit bandwidth rate, and quantization coefficient of the k-th edge device, respectively.

[0028] In one implementation of the first aspect, τ = max k T k ,in z k ,w k p represents the segmentation layer, bandwidth allocation, and transmission power allocation vectors of the k-th edge device, respectively. and Let z represent the segmentation layer of the k-th edge device. k Local computation and upload volume at that time, F k R k (p) and These represent the local computing frequency, unit bandwidth rate, and quantization coefficient of the k-th edge device, respectively.

[0029] Secondly, the present invention provides a federated segmentation learning system based on a non-cellular network, the system comprising multiple edge devices, multiple access points and a cloud server;

[0030] In a non-cellular network, a collaborative service relationship is established between the multiple edge devices and the multiple access points;

[0031] Each edge device is configured with a segmentation layer, bandwidth allocation, and transmission power for the neural network model;

[0032] During the training of the neural network, the neural network model is divided into a local model and a cloud model based on the segmentation layer, and deployed on the multiple edge devices and the cloud server respectively;

[0033] The edge device is used to upload intermediate data output from the segmentation layer in the local model to the cloud server via the collaborative service relationship;

[0034] The cloud server is used to update the global model based on the intermediate data and the cloud model, and to send the updated segmentation layer gradient to the edge device via the collaborative service relationship;

[0035] The edge device is also used to update the local model according to the segmentation layer gradient, and dynamically adjust the segmentation layer, bandwidth allocation and transmission power to perform the next round of training of the neural network model until the neural network model converges.

[0036] As described above, the federated segmentation learning method and system based on cellular-free networks of the present invention have the following beneficial effects:

[0037] (1) With the goal of minimizing the weighted sum of energy consumption and transmission delay, under the constraints of transmission power and transmission bandwidth, a federated segmentation learning framework is adopted, which combines federated segmentation learning and the characteristics of non-cellular networks to achieve resource efficiency and privacy protection of distributed learning.

[0038] (2) By jointly considering access point collaborative services, segmentation layer selection, bandwidth and power resource allocation and quantization transmission strategy, the overall energy consumption and transmission latency are effectively reduced and the efficiency and accuracy of model training are improved while meeting system resource constraints.

[0039] (3) The joint optimization method based on the Alternating Direction Method of Multipliers (ADMM) is adapted to the dynamics and heterogeneity of non-cellular networks. The quantization technique effectively reduces communication overhead while maintaining model accuracy.

[0040] (4) It provides two strategies: joint optimization and step-by-step optimization, which can flexibly adapt to different application scenarios, especially suitable for the general "user-access point-cloud" architecture;

[0041] (5) Compared with traditional federated learning, this invention reduces the computation and communication overhead of edge devices, adapts to the dynamic and heterogeneous nature of non-cellular networks, and provides an efficient solution for edge intelligence. Attached Figure Description

[0042] Figure 1 The diagram shows an architecture of the federated segmentation learning system in one embodiment of the present invention.

[0043] Figure 2 The flowchart shown is a flowchart of an embodiment of the federated segmentation learning method based on non-cellular networks of the present invention;

[0044] Figure 3 The diagram shows a learning flowchart of the federated segmentation learning method based on non-cellular networks according to an embodiment of the present invention.

[0045] Figure 4 The diagram shown is a structural schematic of another embodiment of the federated segmentation learning system based on a non-cellular network according to the present invention. Detailed Implementation

[0046] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0047] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0048] like Figure 1 As shown, the federated segmentation learning system adopts a client-server architecture. On the client side, each terminal device maintains two model components: a client-wide model and a client-local model. The client-wide model is responsible for synchronizing with the server, while the client-local model performs the initial neural network computations. The terminal device transmits intermediate data from the segmentation layer output to the server and receives the returned segmentation layer gradients for local model updates. The server comprises two functional modules: a federated server and a master server. The master server receives intermediate data from each client, performs computations on the remaining neural network layers, and generates global model updates. The federated server aggregates all locally updated models, maintains the client-wide model, and coordinates the entire training process. During federated segmentation learning, the client generates intermediate data through local model computation and uploads it to the master server; after completing subsequent computations, the server sends the segmentation layer gradients back to the client; simultaneously, the server periodically distributes updated global model parameters to each client. This architecture achieves a reasonable division of neural network computation tasks, reducing both the computational burden on terminal devices and the amount of data that needs to be transmitted.

[0049] To address the dual pressures of latency and energy consumption in distributed intelligent applications (such as visual / speech inference on mobile terminals, online analysis and lightweight training of edge video streams, and continuous learning of drones / vehicle sensors) while ensuring the privacy of personal data, this invention presents a federated segmentation learning method based on non-cellular networks. Within a general "edge device—multi-access point—cloud server" architecture, it provides an achievable federated segmentation learning training and optimization process. In a non-cellular collaborative access environment, the edge device only computes the front-end of the model and uploads intermediate data. The cloud server aggregates and updates the global model, minimizing latency and balancing energy consumption for a single round of parallel training while meeting communication and computing resource constraints. In this invention, the edge device can be a smartphone, vehicle-mounted camera, fixed camera, or drone payload (for example only). The upper-layer learning tasks can be classification, retrieval, segmentation, object recognition, anomaly detection, etc. (for example only), and are not limited to specific industry scenarios.

[0050] The technical solutions of the present invention will now be described in detail with reference to the accompanying drawings.

[0051] like Figures 2-3 As shown, in one embodiment, the federated segmentation learning method based on non-cellular networks of the present invention includes steps S1-S6.

[0052] Step S1: In a non-cellular network, establish collaborative service relationships between multiple edge devices and multiple access points.

[0053] Specifically, establishing collaborative service relationships between multiple edge devices and multiple access points includes the following steps:

[0054] 11) For each edge device, establish an uplink with the plurality of access points.

[0055] Among them, the set of edge devices is defined as The set of access points is Edge device k establishes uplinks with several access points l.

[0056] 12) Obtain the large-scale fading coefficient of the uplink corresponding to the edge device and the access point.

[0057] Specifically, based on the uplink, the large-scale fading coefficients of the uplink corresponding to the edge device and the access point are measured, thereby constructing a large-scale fading coefficient matrix. In which element β l.k This indicates the link strength between access point l and edge device k.

[0058] 13) For each edge device, select a preset number of access points in descending order of the large-scale fading coefficient to construct a service set.

[0059] For each edge device k, the large-scale fading coefficient β is used. l.k Select L0 access points from largest to smallest to form a service set.

[0060] 14) When the access point belongs to the service set, it indicates that there is a service relationship with the corresponding edge device; when the access point does not belong to the service set, it indicates that there is no service relationship with the corresponding edge device.

[0061] Here, the service indication matrix D∈{0,1} is defined. L×K If and only if l∈S k Time D l,k =1. In other cases, D l,k =0. That is, l∈S k This indicates the existence of a service relationship with the corresponding edge device. In this invention, collaborative communication between the edge device and the cloud server can be achieved based on the access points in the service set, i.e., multiple access points can collaboratively serve the edge device, thereby constructing a distributed communication network to improve communication efficiency and stability. The dynamic cooperative clustering of the service set is updated periodically / event-triggered (slow time domain) according to user location, occlusion, and slow environmental changes, and is the effective physical layer rate R. k The estimation and stability provide a topological basis, and provide stable physical layer support for subsequent resource allocation.

[0062] Step S2: For each edge device, obtain the segmentation layer, bandwidth allocation, and transmission power of the neural network model.

[0063] Specifically, a set of candidate segmentation layers is determined on the edge device, and the number of candidate segmentation layers on edge device k is set to L. seg And initialize the segmentation vector Initialize bandwidth allocation on the resource side. k ≥0(∑ k w k ≤1) and transmission power The quantization bandwidth is a positive integer, satisfying b∈Z + The segmentation layer candidate vectors are used It means, ∑ l z k,l Let represent the vector when the segmentation layer is l in edge device k. Where is placed within the probability simplex ∑ l z k,l =1,z k,l If the result is greater than or equal to 0, it will be rounded to the nearest one-hot integer. In other words, only the vector corresponding to the segmentation layer is 1 in the segmentation layer vector, and the other vectors are 0.

[0064] Step S3: During the training of the neural network, the neural network model is divided into a local model and a cloud model based on the segmentation layer, and deployed on the multiple edge devices and the cloud server respectively.

[0065] Specifically, edge device k follows the segmentation layer z k The neural network model is divided into a local model and a cloud model. The local model is deployed on the edge device, and the cloud model is deployed on a cloud server. The local model includes a segmentation layer and a layer preceding the segmentation layer in the neural network. The cloud model includes a layer following the segmentation layer in the neural network.

[0066] Step S4: The edge device uploads the intermediate data output from the segmentation layer in the local model to the cloud server via the collaborative service relationship.

[0067] Specifically, during each round of training, the edge device k locally performs training according to the segmentation layer z. k The local computation and uplink upload volumes are as follows:

[0068]

[0069] in, and These represent the local computation and uplink upload volume of edge device k when the segmentation layer is l, respectively.

[0070] To reduce upload overhead, given a quantization bit width *b* for each upload round, this invention employs Lazily Aggregated Quantized Gradient Descent (LAQ) to uniformly quantize the differences in the intermediate data and backfill the residuals to generate quantized data. Thus, the number of bits for uploading the quantized data is approximately *d* *b*, directly mapped to the communication load, where *d* represents the quantization dimension. The difference is represented by *y* = *x* *b*. prev , x, x prev ∈R d These are the parameters for the current round and the previous round, respectively. The quantization radius is taken as R = |y|. ∞ Let the quantization step size be Uniform quantization element by element yields quantized data. Record quantization error In the next round of training, x prev ←x q Achieve residual compensation.

[0071] Step S5: The cloud server updates the global model based on the intermediate data and the cloud model, and sends the updated segmentation layer gradient to the edge device via the collaborative service relationship.

[0072] Specifically, the cloud server executes the cloud model based on the intermediate data, performs federated aggregation, and sends the segmentation layer gradients in the new round of global model to the corresponding edge devices.

[0073] Step S6: The edge device updates the local model according to the gradient of the segmentation layer, and dynamically adjusts the segmentation layer, bandwidth allocation and transmission power to perform the next round of training of the neural network model until the neural network model converges.

[0074] Specifically, the uplink rate is determined by both power and interference. Let the transmit power of edge device k be denoted as . Furthermore, let the transmission power allocation vector for all devices be p = {p1, p2, ..., p...} K The unit bandwidth rate (bit / s / Hz) is modeled as follows: Among them, a k For effective useful link gain, b k For self-interference / mismatch terms, c k For noise and non-cooperative interference, ∑ j≠k d k,j p j d represents the interference power from other edge devices to edge device k. k,j p is the interference intensity coefficient of edge device j to edge device k. j a represents the transmission power of edge device j. k b k c k d k,j Given by a dynamic cooperative clustering strategy and channel state. Bandwidth is allocated according to share w. k ≥0 allocation, satisfying ∑ k w k ≤1, therefore, the latency of one round of edge device k (uplink only) consists of local calculation and upload, as shown below:

[0075]

[0076] Among them, z k ,w k p represents the segmentation layer, bandwidth allocation, and sum of the k-th edge device, respectively. and Let z represent the segmentation layer of the k-th edge device. k Local computation and upload volume at that time, F k R k (p) and Let represent the local computation frequency, unit bandwidth rate, and quantization coefficient of the k-th edge device, respectively. The global latency of a single round of parallel synchronous training across multiple edge devices is determined by the slowest edge device. Therefore, the global transmission latency τ = max k T k .

[0077] Regarding energy consumption, this invention employs a linear-affine power consumption model, where the energy required for one training round of edge device k is... Among them κ k Calculate the energy consumption constant in units of 1.

[0078] The edge device updates its local model based on the segmentation layer gradient and proceeds to the next training round until a preset number of training rounds is reached or the target value λτ+μ is less than a preset threshold, where τ represents the global transmission latency, and E... k Let λ represent the energy consumption of the k-th edge device, and λ and μ represent the tradeoff coefficients. Simultaneously, at short time intervals, based on real-time network environment information (such as location and channel status), the segmentation layer selection, bandwidth allocation, transmission power, and quantization level are dynamically adjusted to optimize real-time response speed and maintain model training accuracy.

[0079] In this invention, the above target value can be equivalently linearized as follows:

[0080]

[0081] To achieve efficient solutions, this invention employs a distributed alternating optimization framework, which decomposes the problem into multiple parallel subproblems and solves them iteratively in alternation.

[0082] First, the segmentation layer selection and bandwidth allocation are updated alternately using ADMM to achieve consensus updates of variables on the user side and the central side, thus balancing computational overhead. The solution is then performed in parallel on the device side.

[0083]

[0084] Where ρ>0 is the augmented Lagrange penalty parameter of ADMM, u k For corresponding consistency constraints The scaling dual variable (Lagrange multiplier), z k The subproblem is linear selection (selecting the segmentation layer that minimizes the cost, adding slight sharpening regularization to promote one-hot encoding if necessary), w k The subproblem is a one-dimensional convex problem:

[0085]

[0086] A closed-loop cubic / Newton iteration can be used for a fast solution, setting w = {w1, ..., w}. K} represents the bandwidth share vector for all devices, projected onto the center side. in To reach the feasible set The Euclidean projection operator is used to satisfy the total bandwidth constraint and update the dual variable u accordingly.

[0087] Secondly, the power allocation uses the generalized Dinkelbach method to maximize the worst-case energy efficiency. To achieve a trade-off between energy consumption and time delay. Solve the parameterized subproblem max under given parameter Λ. p min k {R k (p)-Λ(ζ0P u p k +ζ1)}. The implementation uses Sequential Least Squares Programming (SLSQP) to numerically solve for the problem under box constraints (and optional linear disturbance / power budget constraints), and uses... Update until min k {R k -ΛG k The absolute value of} is lower than the set threshold.

[0088] Finally, delayed aggregation quantization is performed independently on the edge device side in each round, requiring only the exchange of quantized vectors / codewords with the cloud. Furthermore, the quantization bit width *b* and quantization dimension *d* are directly mapped to the communication bits *d·b*, using compression coefficients. Feedback to the latency model employs an alternating strategy of "fixed p first ADMM(z, w) → fixed z, w then Dinkelbach(p)", which typically achieves stable convergence after several rounds; the service set independently updates and refreshes the rate parameter R in the slow time domain. k .

[0089] By alternately executing the above methods, this invention can obtain a convergent joint optimal solution under limited resource constraints. Through the above methods, this invention can implement a resource dynamic optimization strategy based on federated segmentation learning in cellular-free networks. It significantly improves the computing efficiency and communication performance of edge devices, meets the distributed learning needs in resource-constrained environments, and avoids the energy and latency imbalance problem in model training through joint optimization strategies.

[0090] It should be noted that, in order to adapt to different applications, the operating points of this invention (such as L0, λ / μ, bit width b, fairness / maximum delay constraints) can be adjusted according to task-side indicators (such as end-to-end 95% latency, device power limit, task accuracy limit); if it is necessary to further improve robustness, adaptive bit width, block sparsity, projection fairness regularization, etc. can be replaced or superimposed. These variations do not change the core process and distributed solution structure of this invention.

[0091] The scope of protection of the federated segmentation learning method based on non-cellular networks described in this embodiment is not limited to the execution order of the steps listed in this embodiment. Any solution implemented by adding, subtracting, or replacing steps in the prior art based on the principles of this invention is included within the scope of protection of this invention.

[0092] This invention also provides a federated segmentation learning system based on a cellular network. The federated segmentation learning system based on a cellular network can implement the federated segmentation learning method based on a cellular network described in this invention. However, the implementation device of the federated segmentation learning system based on a cellular network described in this invention includes, but is not limited to, the structure of the federated segmentation learning system based on a cellular network listed in this embodiment. All structural modifications and substitutions of the prior art made in accordance with the principles of this invention are included within the protection scope of this invention.

[0093] like Figure 4 As shown, in one embodiment, the federated segmentation learning system based on a non-cellular network of the present invention includes multiple edge devices 41, multiple access points 42, and a cloud server 43.

[0094] In a non-cellular network, a collaborative service relationship is established between the plurality of edge devices 41 and the plurality of access points 42.

[0095] Each edge device 41 is equipped with a segmentation layer, bandwidth allocation, and transmission power for the neural network model.

[0096] During the training of the neural network, the neural network model is divided into a local model and a cloud model based on the segmentation layer, and deployed on the multiple edge devices 41 and the cloud server 43 respectively.

[0097] The edge device 41 is used to upload the intermediate data output by the segmentation layer in the local model to the cloud server via the collaborative service relationship.

[0098] The cloud server 43 is used to update the global model based on the intermediate data and the cloud model, and to send the updated segmentation layer gradient to the edge device via the collaborative service relationship;

[0099] The edge device 41 is also used to update the local model according to the segmentation layer gradient, and dynamically adjust the segmentation layer, bandwidth allocation and transmission power to perform the next round of training of the neural network model until the neural network model converges.

[0100] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, or methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or modules or units may be electrical, mechanical, or other forms.

[0101] The modules / units described as separate components may or may not be physically separate. The components shown as modules / units may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules / units can be selected to achieve the objectives of the embodiments of the present invention, depending on actual needs. For example, the functional modules / units in the various embodiments of the present invention may be integrated into one processing module, or each module / unit may exist physically separately, or two or more modules / units may be integrated into one module / unit.

[0102] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0103] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A federated split learning method based on a cell-free network, characterized in that, The method comprises the following steps: In a cell-free network, a cooperative service relationship between a plurality of edge devices and a plurality of access points is established; For each edge device, a split layer of a neural network model, a bandwidth allocation and a transmission power are obtained; During the neural network training, the neural network model is divided into a local model and a cloud model based on the split layer, and is respectively deployed on the plurality of edge devices and a cloud server; The edge device uploads intermediate data output by the split layer in the local model to the cloud server via the cooperative service relationship; The cloud server updates a global model based on the intermediate data and the cloud model, and issues updated split layer gradients to the edge device via the cooperative service relationship; The edge device updates the local model according to the split layer gradients, and dynamically adjusts the split layer, the bandwidth allocation and the transmission power to perform the next round of training of the neural network model until the neural network model converges.

2. The federated split learning method based on the cell-free network according to claim 1, wherein: Establishing a cooperative service relationship between a plurality of edge devices and a plurality of access points comprises the following steps: For each edge device, an uplink is established with the plurality of access points; Large-scale fading coefficients of the uplink corresponding to the edge device and the access point are obtained; For each edge device, a preset number of access points are selected according to the large-scale fading coefficients from large to small to construct a service set; When the access point belongs to the service set, it indicates that there is a service relationship with the corresponding edge device; when the access point does not belong to the service set, it indicates that there is no service relationship with the corresponding edge device.

3. The federated split learning method based on the cell-free network according to claim 1, wherein: The edge device uploads intermediate data output by the split layer in the local model to the cloud server via the cooperative service relationship, comprising the following steps: The edge device uses delay aggregation quantization to uniformly quantize the difference of the intermediate data and generates quantization data through residual backfilling; The quantization data is uploaded to the cloud server via the cooperative service relationship.

4. The federated split learning method based on the cell-free network according to claim 3, characterized in that: The number of bits of the quantization data is d·b, where d represents the quantization dimension and b represents the quantization bit width.

5. The federated split learning method based on the cell-free network according to claim 1, wherein: The split layer and the bandwidth allocation are adjusted based on the alternating direction multiplier method.

6. The federated split learning method based on the cell-free network according to claim 1, wherein: The transmission power is adjusted based on the generalized Dinkelbach method.

7. The federated split learning method based on the cell-free network according to claim 1, wherein: When any of the following conditions is met, the neural network model converges: 1) reaching a preset training round number; 2) target value λτ+μ∑ k E k less than a preset threshold, where τ represents a global transmission delay, E k represents the energy consumption of the kth edge device, and λ and μ represent trade-off coefficients.

8. The federated split learning method based on the cell-free network according to claim 7, characterized in that: where z k ,w k ,p represent the partition layer, bandwidth allocation and transmission power allocation vector of the kth edge device, respectively, and represent the local computation and upload amount of the kth edge device when the partition layer is z k , k 、 R k (q) and represent the unit computation energy consumption constant, radio frequency transmission power constant, unit bandwidth rate and quantization coefficient of the kth edge device, respectively.

9. The federated split learning method based on the cell-free network according to claim 7, characterized in that: T = max k T k where z k ,w k ,prespectively represent the partition layer, bandwidth allocation and transmission power allocation vector of the kth edge device, and respectively represent the local computation and upload amount of the kth edge device when the partition layer is z k F k ,R k (p) and respectively represent the local computation frequency, unit bandwidth rate and quantization coefficient of the kth edge device. 10.A federated split learning system based on a cell-free network, characterized in that, The system comprises a plurality of edge devices, a plurality of access points and a cloud server; In a cell-free network, a cooperative service relationship is established between the plurality of edge devices and the plurality of access points; Each edge device is provided with a split layer of a neural network model, a bandwidth allocation and a transmission power; During the neural network training, the neural network model is divided into a local model and a cloud model based on the split layer, and is respectively deployed on the plurality of edge devices and a cloud server; The edge device is configured to upload intermediate data output by the split layer in the local model to the cloud server via the cooperative service relationship; The cloud server is configured to update a global model based on the intermediate data and the cloud model, and to issue updated segmentation layer gradients to the edge device via the collaborative service relationship. The edge device is further configured to update the local model according to the segmentation layer gradients, and to dynamically adjust a segmentation layer, a bandwidth allocation and a transmission power for training the neural network model in the next round until the neural network model converges.

Images