A Method and System for Reducing the Energy Consumption of a Wireless Computing Power Network Based on Federated Learning

Abstract

The present invention provides a method and system for reducing the energy consumption of a wireless computing power network based on federated learning. The method is characterized in that it includes the following steps: step 1, obtaining application requirements corresponding to the application plane in the wireless computing power network; step 2, making a computing request based on the obtained application requirements; step 3, carrying an initialized federated learning model on each computing node in the computing plane in the wireless computing power network, and determining the structural parameters of the initialized federated learning model; step 4, using the computing request obtained in step 2 to train the initialized federated learning model carried in step 3 to obtain a trained federated learning model; step 5, using the trained federated learning model to perform business processing on received user data. The present invention can accelerate the convergence speed of federated learning in WCPN and reduce the energy consumption of traditional asynchronous federated learning.

Description

Technical Field

[0001] The present invention belongs to the field of network communications, and specifically relates to a method and system for reducing energy consumption of wireless computing networks based on federated learning. Background Art

[0002] In the 6G era, wireless networks will generate massive amounts of data and information. Over 50% of this data will need to be analyzed, processed, and stored at the network edge. This requires significant communication resources and computing power, posing unprecedented challenges to current computing frameworks. Mobile Edge Computing (MEC) primarily relies on the capabilities of individual edge nodes. Excessive or unbalanced requests can lead to node overload or communication congestion. Coordinating the network, computing, storage, and algorithm resources of edge nodes to collaboratively process this massive data is imperative, ultimately enabling ubiquitous computing and intelligent computing.

[0003] Most current computing network frameworks integrate edge computing and cloud computing, such as the compute-first network framework, which is a computing platform composed of computing nodes such as terminals, edge nodes, and cloud nodes. However, this framework does not implement intelligent network services and cannot propose dynamic task strategies. In addition, research has found that integrating federated learning can effectively allocate resources, improve the efficiency of computing task completion, and achieve energy conservation. However, existing technologies ignore the significant impact of the neural network structure of computing nodes in federated learning on overall energy consumption. In fact, the depth of the neural network of each node not only affects the amount of data transmitted, but also the performance of the collaborative learning model. In summary, existing technologies do not consider intelligent network services and the nodes' adjustable computing and dynamic network strategies for specific tasks. This makes it impossible to optimize energy consumption, which in turn reduces the computing power of 6G networks. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for reducing the energy consumption of wireless computing power networks based on federated learning, which solves the problem that energy consumption in wireless computing power networks in the prior art cannot be optimized, thereby reducing the computing power of 6G networks.

[0005] In order to achieve the above object, the technical solution adopted in the present invention is:

[0006] The present invention provides a method for reducing energy consumption of a wireless computing network based on federated learning, which is characterized by comprising the following steps:

[0007] Step 1: Obtain application requirements corresponding to the application plane in the wireless computing network;

[0008] Step 2: Propose a computing request based on the obtained application requirements;

[0009] Step 3: Initialize the federated learning model on each computing node in the computing plane of the wireless computing network and determine the structural parameters of the initialized federated learning model;

[0010] Step 4: Use the computation request obtained in step 2 to train the initialized federated learning model loaded in step 3 to obtain a trained federated learning model.

[0011] Step 5: Use the trained federated learning model to perform business processing on the received user data.

[0012] Preferably, in step 3, the federated learning model is initialized and the structural parameters of the initialized federated learning model are determined. The specific method is:

[0013] Set the expression for the energy consumption of all computing nodes for one local model training;

[0014] Set the expression for the energy consumption of all computing nodes performing one uplink communication;

[0015] Based on the expression of energy consumption for one local model training and the expression of energy consumption for one uplink communication, the expression of total energy consumption of all computing nodes is obtained;

[0016] Setting expressions for the representation capability of the neural network and expressions for the test performance of the neural network;

[0017] Based on the expression of the total energy consumption of all computing nodes, the expression of the representation ability of the neural network and the expression of the test performance of the neural network, the total energy consumption optimization problem is obtained;

[0018] The obtained total energy consumption optimization problem is solved to obtain the structural parameters of the initialized federated learning model.

[0019] Preferably, in step 4, the calculation request obtained in step 2 is used to train the initialized federated learning model carried out in step 3 to obtain a trained federated learning model. The specific method is:

[0020] S401, each computing node uses the local corresponding data set as the input to initialize the federated learning model and obtain the output of the local model;

[0021] S402, performing global aggregation processing on the local model outputs corresponding to all computing nodes to obtain a global aggregate model;

[0022] S403, distributing the output of the global aggregation model to the local model corresponding to each computing node, and updating the local model;

[0023] S404, each computing node uses the corresponding local data set as the input of the updated local model to obtain the output of the local model;

[0024] S405, repeat S402 to S404 until the model converges to obtain a trained federated learning model.

[0025] Preferably, in S402, the local model outputs corresponding to all computing nodes are globally aggregated to obtain a global aggregate model. The specific method is:

[0026] An energy-saving asynchronous aggregation algorithm is used to globally aggregate the local model outputs corresponding to all computing nodes.

[0027] Preferably, an energy-saving asynchronous aggregation algorithm is used to perform global aggregation processing on the local model outputs corresponding to all computing nodes. The specific method is:

[0028] S4021, divide all computing nodes into multiple clusters based on the local model output;

[0029] S4022, determining the coordination time between multiple clusters;

[0030] S4023, determining collaborative learning between all corresponding computing nodes in each cluster;

[0031] S4024, performing internal aggregation within each cluster based on the determined collaborative learning to obtain an intra-cluster aggregation model;

[0032] S4025, performing aggregation among all clusters according to the determined collaboration time;

[0033] S4026, iteratively execute S4023 to S4025 until the loss function converges, thereby obtaining a global aggregation model.

[0034] Preferably, in S4022, the collaboration time between the multiple clusters is determined by:

[0035] In order to determine the collaboration time of all computing nodes, the inter-cluster collaboration index χ is set on the aggregation node. The update rule of χ is as follows:

[0036]

[0037] Among them, τ i is the number of local training rounds adopted by computing node i; C Indicates the communication counter of cluster c.

[0038] Preferably, in S4023, collaborative learning between all corresponding computing nodes in each cluster is determined, specifically by:

[0039] In each cluster, the computing nodes use the optimal local model training parameters for learning. The rules for local model training updates are:

[0040]

[0041] Among them, w i is the parameter obtained by the calculation node through the local model; η≥0 represents the step size, is the loss function F i (w i ) gradient.

[0042] A system for reducing energy consumption of wireless computing networks based on federated learning, comprising:

[0043] A demand acquisition unit, configured to acquire application requirements corresponding to the application plane in the wireless computing network;

[0044] A task proposing unit is used to propose computing requests based on the obtained application requirements;

[0045] A model loading unit is used to load the initialized federated learning model on each computing node in the computing plane in the wireless computing network and determine the structural parameters of the initialized federated learning model;

[0046] The model optimization unit is used to train the initialized federated learning model using the obtained computing request to obtain a trained federated learning model;

[0047] The model application unit is used to use the trained federated learning model to perform business processing on the received user data.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The present invention provides a method for reducing energy consumption in wireless computing networks based on federated learning. Unlike conventional wireless communication networks, the present invention jointly optimizes the computing strategies and collaborative learning strategies of each computing node to minimize the total energy consumption of all computing nodes. Specifically:

[0050] First: Dynamic computing strategy adjustment: The present invention can dynamically adjust the computing tasks of each computing node according to specific computing task requirements, thereby reducing the computing pressure of edge computing nodes.

[0051] Second: Intelligent Services. This paper designs a task- and resource-aware federated learning model (FedTAR), which minimizes total energy consumption by jointly optimizing the computing strategies of individual computing nodes and their collaborative learning strategies according to specific task requirements.

[0052] Third: Based on the solution of the optimization problem, the neural network depth of the computing nodes and the collaboration frequency between nodes can be adjusted according to the specific computing task requirements and resource constraints, further finding the minimum total energy consumption of the model.

[0053] Fourth: In order to further adapt to heterogeneous computing nodes, an energy-saving FedTAR asynchronous aggregation algorithm is proposed, which accelerates the convergence speed of federated learning in WCPN and reduces the energy consumption of traditional asynchronous federated learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is the Wireless Computing Power Network (WCPN).

[0055] Figure 2 It is a federated learning model between computing nodes.

[0056] Figure 3 The relationship between the energy consumption of the computing node and the local training round τ and the depth of the neural network.

[0057] Figure 4 is a graph showing the relationship between the exact value of the local training round τ and the ratio ξ (the ratio of communication and computation energy consumption).

[0058] Figure 5 It is a federated learning model based on asynchronous aggregation.

[0059] Figure 6 The relationship between energy consumption and local training round τ for different ratio intervals ξ.

[0060] Figure 7 Comparison of training losses for different schemes.

[0061] Figure 8 Comparison of model accuracy in different schemes.

[0062] Figure 9 Comparison of energy consumption of computing nodes under different schemes. DETAILED DESCRIPTION

[0063] The present invention will be described in further detail below with reference to the accompanying drawings.

[0064] This paper designs a task- and resource-aware federated learning model, called FedTAR, in a wireless computing power network (WCPN). By jointly optimizing the computing strategies and collaborative learning strategies of each computing node, the total energy consumption of all computing nodes is minimized. Based on the solution of the optimization problem, the neural network depth of the computing nodes and the collaboration frequency between nodes can be adjusted according to the specific computing task requirements and resource constraints. To further adapt to heterogeneous computing nodes, this technology proposes an energy-saving FedTAR asynchronous aggregation algorithm, which accelerates the convergence of federated learning in WCPN. Numerical results show that this scheme outperforms existing work in terms of learning accuracy, convergence speed, and energy saving.

[0065] The present invention includes the following technical details, and the specific operation mode is as follows:

[0066] 1. The task requirements under the WCPN framework (including the computing power plane and application layer) are based on the computational process of federated learning.

[0067] The present invention establishes a WCPN framework network, in which the computing power plane is composed of 100 different computing nodes, such as terminals, edge computing servers and cloud servers. At the same time, the computing nodes with higher configuration are selected from the 100 different computing nodes to serve as aggregation nodes; Figure 1 shown.

[0068] Step 1, data collection: The application plane collects application requirements corresponding to smart industry, smart transportation, smart home and smart agriculture.

[0069] Step 2, data conversion: Propose a computing request based on the obtained application requirements; the obtained computing request forms a data set;

[0070] Step 3, data submission: The proposed computing request is automatically routed to the computing node in the computing power plane.

[0071] Step 4: The computing node trains the computing request locally to obtain a local model.

[0072] Calculate the node i in the dataset Δ i The loss function on is expressed as:

[0073]

[0074] Among them, D i Represents the dataset Δ i The size of w i Indicates the model parameters currently trained by computing node i, J i (w i ) represents the loss function of the training model. Calculate the dataset Δ of node i i A set of input and output pairs Given, x j ∈R d Represents an input sample vector with d features, y j ∈R represents sample x j The markup output value of .

[0075] The training model parameters of the present invention use a gradient descent algorithm with a learning rate of 0.01, which can be expressed as:

[0076]

[0077] Among them, η≥0 represents the step size, is the loss function F i (w i ) gradient.

[0078] Step 5: Each computing node uploads the local model trained in step 4 to the aggregation node for global aggregation. The global aggregation of the model can be expressed as:

[0079]

[0080] Among them, w i is the parameter obtained by the computing node through the local model, is the total data volume of all computing nodes.

[0081] Step 6: Update the global model and return the updated global model from the aggregation node to each computing node. The global loss function can be expressed as:

[0082]

[0083] Here, w is a global model parameter.

[0084] The global model parameters are trained using the gradient descent algorithm.

[0085] The general upper limit of the global model parameters during the iteration process can be expressed as:

[0086]

[0087] where O(·) is a measure of polynomial complexity and Γ(ε,θ) is affected by the global accuracy threshold ε and the local accuracy threshold θ. When ε and θ are small, the learning process requires more global and local updates.

[0088] Step 7: Repeat steps 4, 5, and 6 for τ rounds until the model converges, that is, the local accuracy reaches the threshold θ (0≤θ≤1) and the global model accuracy approaches the threshold ε (0≤ε≤1). The process of computing the federated learning model between nodes is as follows: Figure 2 As shown, the relationship between local training rounds and energy consumption is as follows Figure 3shown.

[0089] Step 8, data feedback: finally return the result corresponding to the application computing requirements.

[0090] 2. Initialize the global model: optimize the model structure to achieve the goal of saving energy.

[0091] Step 1: The energy consumed by the computing node for one local model training can be expressed as:

[0092]

[0093] Among them, v i represents the number of CPU cycles required to train a single data sample, f i represents the CPU frequency of computing node i, k i / 2 represents the effective capacitance coefficient of the computing chipset of computing node i, D i is the size of the dataset on computation node i.

[0094] Step 2: Calculate the energy consumption of all computing nodes performing one local training, which can be expressed as:

[0095]

[0096] Step 3: The neural network has the best performance when the number of neurons in each hidden layer is the same. The number of neurons in the hidden layer can be calculated as follows:

[0097]

[0098] Among them, m in is the number of input neurons, m out is the number of output neurons and l is the depth of the neural network.

[0099] Step 4: Combined with step 3, the number of model parameters in the fully connected neural network can be expressed as:

[0100]

[0101] Where s is the number of neurons in the hidden layer.

[0102] Step 5: Calculate the energy consumed by node i in each uplink communication, which can be expressed as:

[0103]

[0104] Among them, b is the size of a model parameter in the neural network and is a constant. i is the transmission power of computing node i, B i is the transmission bandwidth of computing node i, hi is the channel gain of computing node i, I i is the communication background noise of computing node i.

[0105] Step 6: Calculate the energy consumption of one uplink communication of all computing nodes, which can be expressed as:

[0106]

[0107] Step 7: Calculate the total energy consumption, combined with the frequency τ of federated learning and the neural network depth l, which can be expressed as:

[0108] E glob (τ,l)=Γ(ε,θ)·(τ·E cmp +E com ) (15)

[0109] Where Γ(ε,θ) is the upper limit of global iteration, τ is the update frequency of model learning, and l is the depth of the neural network. The relationship between the energy consumption of the computing node and the local training round τ and the depth of the neural network l is as follows: Figure 3 As shown, the relationship between the exact value τ of the local training round and the ratio ξ (the ratio of communication and computation energy consumption) is as follows Figure 4 shown.

[0110] Step 8: The representation ability of a neural network is a measure of its ability to fit the training data, which can be represented by the Betti number. The range of Betti can be expressed as:

[0111]

[0112] Where l is the depth of the neural network, r is the degree of the activation function used in the neural network, and m in is the number of input neurons in the neural network.

[0113] Step 9: Test performance is to measure the gap between the neural network training results and the actual situation, which can be represented by Rademacher complexity. Rademacher complexity can be expressed as:

[0114]

[0115] in, γ is the coefficient of the regularization term, α and β are constants used to calculate the norm, ranging from [1, 2], and d is the input data set x j The dimension of μ is x j size.

[0116] Step 10: Optimize the total energy consumption by combining the representation capability of the task training model, test performance, and training accuracy constraints. This can be expressed as:

[0117]

[0118]

[0119]

[0120] e -τ ≤1 (16c)

[0121] Among them, constraint (16a) guarantees the test performance of the neural network and requires that the test error is less than δ, and constraint (16b) guarantees the representation ability of the neural network and requires that the lower limit of the representation ability is greater than Constraint (16c) ensures that the collaboration between nodes can achieve the training objectives based on federated learning.

[0122] The computational strategies of the nodes (e.g., the depth of the neural network, l) and their coordination methods (e.g., their update frequency, τ) are jointly optimized to minimize the total energy consumption. In particular, the optimization problem determines the optimal neural network depth that minimizes the energy consumption of the computational nodes for a specific computational task. The update frequency between the computational nodes is then optimized to minimize energy consumption while achieving the training goal. By utilizing this optimization scheme, the computational strategies of the nodes can be dynamically adjusted according to the requirements and complexity of the task.

[0123] 3. FedTAR energy-saving asynchronous aggregation algorithm, to achieve the global aggregation of step 5 in 1 and reduce energy consumption, the model is as follows Figure 5 shown.

[0124] Step 1: Initialize the computing nodes. First, the aggregation node determines the structure of the training model based on the required representation power and test performance of the computing task. Then, the initialized model is deployed on each computing node. The ratio of communication and computing consumption of computing node i is given by Finally, each computing node sends the ratio ξ to the aggregation node i .

[0125] Step 2: Clustering of computing nodes. Clustering nodes divide computing nodes into C clusters according to ξ (the present invention assumes that clusters are sorted in ascending order according to ξ). The number of computing nodes in cluster c (c = 1, 2, ..., C) is expressed as N c , denote the total data size of cluster c as D c In addition, a communication counter λ is set for each cluster c = 0, record the number of intra-cluster aggregations. For each intra-cluster aggregation, the counter is incremented by 1.

[0126] Step 3: Determine the collaboration time. To determine the collaboration time of all computing nodes, set the inter-cluster collaboration index χ on the aggregation node. The update rule of χ is as follows:

[0127]

[0128] Among them, τ i is the number of local training rounds used by computing node i; N is the number of computing nodes; λ C The communication counter for cluster c is used to record the number of intra-cluster aggregations. This rule prevents excessively long intervals between computations between adjacent inter-cluster aggregations, thereby accelerating convergence. When the indicator x = true, the latest model parameters for each cluster are aggregated at the aggregation node for inter-cluster aggregation. Otherwise, the compute node performs local training or intra-cluster aggregation.

[0129] This rule avoids excessively long intervals between computations between adjacent clusters, thereby accelerating convergence. When the indicator x = true, the latest model parameters for each cluster are aggregated at the aggregation node for inter-cluster aggregation. Otherwise, the compute node performs local training or intra-cluster aggregation.

[0130] Step 4: Collaborative learning between computing nodes. In each cluster, computing nodes use the optimal local model training parameters for learning. The rules for local model training updates are:

[0131]

[0132] Step 5: Intra-cluster aggregation. The computing node sends the latest local model parameters to the aggregation node for intra-cluster aggregation, and then broadcasts the updated model parameters to all computing nodes in the cluster, which can be expressed as:

[0133]

[0134] Among them, w c represents the intra-cluster aggregation model of cluster C, represents the model parameter of the last intra-cluster aggregation, and h∈[0,1] represents the coordination control coefficient, which is used to adjust the collaboration between clusters.

[0135] Step 6: Inter-cluster aggregation. When the indicator x = true, the latest local model parameters of all clusters are aggregated at the aggregation node. The aggregation node then performs inter-cluster aggregation and broadcasts the new model parameters w to all compute nodes.

[0136] Step 7: Steps 4, 5, and 6 are iteratively performed until the loss function converges or the target accuracy is reached.

[0137] The present invention also provides a system for reducing energy consumption of a wireless computing network based on federated learning, comprising:

[0138] A demand acquisition unit, configured to acquire application requirements corresponding to the application plane in the wireless computing network;

[0139] A task proposing unit is used to propose computing requests based on the obtained application requirements;

[0140] A model loading unit is used to load the initialized federated learning model on each computing node in the computing plane in the wireless computing network and determine the structural parameters of the initialized federated learning model;

[0141] The model optimization unit is used to train the initialized federated learning model using the obtained computing request to obtain a trained federated learning model;

[0142] The model application unit is used to use the trained federated learning model to perform business processing on the received user data.

[0143] Model comparison:

[0144] The Fixed and Adaptive schemes are used as benchmarks. In the Fixed scheme, the number of local training rounds of a compute node is fixed during federated learning; the Adaptive scheme adaptively adjusts the number of local training rounds based on a limited resource budget. Finally, the proposed scheme, called FedTAR, is compared with the two aforementioned schemes in terms of learning accuracy, convergence speed, and energy conservation.

[0145] Figure 6 Describes the energy consumption of computing nodes in FedTAR under different ratios ξ (ratio of communication to computing energy consumption) and different local training rounds. As can be seen from the figure, the optimal local training round is Figure 6 τ in (b) * =2 and ξ∈(1.720,8.019], in Figure 6 In (d), ξ∈(27.878,100], the optimal number of local training rounds is 4. This is because Figure 6 Low values of ξ in (b) indicate that computation consumes more energy than communication, so the computation nodes need to be tuned to perform frequent updates (corresponding to low τ * ), thus achieving energy efficiency. Frequent local updates between two global aggregates (i.e., τ = 3 or τ = 4) will lead to underutilization of communication capacity, resulting in additional energy consumption. Numerical results show that the proposed dynamic learning scheme can adapt to the time-varying communication and computation capabilities of computing nodes, and the ratio ξ can capture the capabilities of computing nodes and is therefore suitable for adjusting the update strategy.

[0146] Figure 7The training loss of the FedTAR scheme and the baseline schemes (Fixed and Adaptive) is compared as a function of the number of global communication rounds. As can be seen in the figure, the training loss of all three schemes decreases as the number of global communication rounds increases. After 10 rounds, the training losses of the three schemes begin to converge. This demonstrates that the FedTAR scheme effectively guarantees the convergence of federated learning. Note that the training loss of the FedTAR scheme converges faster than the other two schemes and the baseline. This is because the local training rounds of the compute nodes in the FedTAR scheme are dynamically adjusted based on the capabilities of the compute nodes. In the Fixed scheme, heterogeneous compute nodes use the same update strategy, resulting in a straggler effect. The Adaptive scheme adjusts its update strategy based on the total remaining resources, rather than the capabilities of each compute node. Therefore, the capabilities of each compute node cannot be fully utilized, resulting in slower convergence than the FedTAR scheme.

[0147] Figure 8 The figure shows how model accuracy changes with the number of global communication rounds. As can be seen, the accuracy improves significantly initially, then gradually converges with increasing global communication rounds. Compared to the other two baselines, the FedTAR scheme achieves the highest accuracy with fewer global rounds because, in this case, FedTAR can fully utilize the computational and communication capabilities of heterogeneous computing nodes.

[0148] Figure 9 The energy consumption of the three schemes under different numbers of clusters is compared. As can be seen in the figure, the energy consumption of the FedTAR scheme is generally lower than that of the other two benchmarks. The FedTAR scheme can effectively reduce energy consumption because all computing nodes adopt the optimal local training rounds to participate in learning. In addition, the FedTAR scheme can effectively eliminate the lagging effect of cluster nodes in communication and computing capabilities (e.g., transmission power, CPU frequency, channel gain). In contrast, in the Fixed scheme, all computing nodes adopt a fixed local training round, resulting in additional local training for some capable nodes, which in turn leads to more energy consumption. For the Adaptive scheme, the computing strategy is adaptively adjusted according to the resource status of the computing nodes. Then, the same local training rounds are deployed on all computing nodes, similar to the 12 rounds of the Fixed scheme, which leads to additional energy consumption due to excessive local training or global communication.

Claims

1. A method for reducing energy consumption of wireless computing networks based on federated learning, characterized in that: The following steps are involved: Step 1: Obtain application requirements corresponding to the application plane in the wireless computing network; Step 2: Propose a computing request based on the obtained application requirements; Step 3: Initialize the federated learning model on each computing node in the computing plane of the wireless computing network and determine the structural parameters of the initialized federated learning model; Step 4: Use the computation request obtained in step 2 to train the initialized federated learning model loaded in step 3 to obtain a trained federated learning model. Step 5: Use the trained federated learning model to process the received user data. In step 3, initialize the federated learning model and determine the structural parameters of the initialized federated learning model. The specific method is: Set the expression for the energy consumption of all computing nodes for one local model training; Set the expression for the energy consumption of all computing nodes performing one uplink communication; Based on the expression of energy consumption for one local model training and the expression of energy consumption for one uplink communication, the expression of total energy consumption of all computing nodes is obtained; Setting expressions for the representation capability of the neural network and expressions for the test performance of the neural network; Based on the expression of the total energy consumption of all computing nodes, the expression of the representation ability of the neural network and the expression of the test performance of the neural network, the total energy consumption optimization problem is obtained; Solve the obtained total energy consumption optimization problem to obtain the structural parameters of the initialized federated learning model; In step 4, the computation request obtained in step 2 is used to train the initialized federated learning model loaded in step 3 to obtain a trained federated learning model. The specific method is: S401, each computing node uses the local corresponding data set as the input to initialize the federated learning model and obtain the output of the local model; S402, performing global aggregation processing on the local model outputs corresponding to all computing nodes to obtain a global aggregate model; S403, distributing the output of the global aggregation model to the local model corresponding to each computing node, and updating the local model; S404, each computing node uses the corresponding local data set as the input of the updated local model to obtain the output of the local model; S405, repeat S402 to S404 until the model converges to obtain a trained federated learning model.

2. The method for reducing energy consumption of wireless computing network based on federated learning according to claim 1, characterized in that: S402: Globally aggregate the local model outputs corresponding to all computing nodes to obtain a global aggregate model. The specific method is: An energy-saving asynchronous aggregation algorithm is used to globally aggregate the local model outputs corresponding to all computing nodes.

3. The method for reducing energy consumption of wireless computing network based on federated learning according to claim 2, characterized in that: The energy-saving asynchronous aggregation algorithm is used to globally aggregate the local model outputs corresponding to all computing nodes. The specific method is: S4021, divide all computing nodes into multiple clusters based on the local model output; S4022, determining the coordination time between multiple clusters; S4023, determining collaborative learning between all corresponding computing nodes in each cluster; S4024, performing internal aggregation within each cluster based on the determined collaborative learning to obtain an intra-cluster aggregation model; S4025, performing aggregation among all clusters according to the determined collaboration time; S4026, iteratively execute S4023 to S4025 until the loss function converges, thereby obtaining a global aggregation model.

4. The method for reducing energy consumption of wireless computing network based on federated learning according to claim 3, characterized in that: In S4022, the coordination time between the multiple clusters is determined. Specifically, the method is: In order to determine the coordination time of all computing nodes, set the inter-cluster coordination indicator on the aggregation node , The update rules are as follows: (17) in, Is a computing node The number of local training rounds used; Indicates the communication counter of cluster c.

5. The method for reducing energy consumption of wireless computing network based on federated learning according to claim 3, characterized in that: In S4023, collaborative learning between all corresponding computing nodes in each cluster is determined. The specific method is: In each cluster, the computing nodes use the optimal local model training parameters for learning. The rules for local model training and update are: in, It is the parameter obtained by the computing node through the local model; represents the step length, is the loss function gradient.

6. A system for reducing energy consumption of wireless computing networks based on federated learning, characterized in that: The method according to claim 1, comprising: A demand acquisition unit, configured to acquire application requirements corresponding to the application plane in the wireless computing network; A task proposing unit is used to propose computing requests based on the obtained application requirements; A model loading unit is used to load the initialized federated learning model on each computing node in the computing plane in the wireless computing network and determine the structural parameters of the initialized federated learning model; The model optimization unit is used to train the initialized federated learning model using the obtained computing request to obtain a trained federated learning model; The model application unit is used to use the trained federated learning model to perform business processing on the received user data.

Images