A heterogeneous wireless federation learning utility perception scheduling and bandwidth allocation method

Abstract

The present application relates to a kind of heterogeneous wireless federal learning utility perception scheduling and bandwidth allocation method, comprising: by edge client reporting lightweight metadata, parameter server calculates the comprehensive utility value of data scale utility, data diversity utility and dynamic gradient projection utility weighted combination, according to sorting selection participating client;For selected client, with the minimum utility weighted communication delay as target, bandwidth resource optimization model is established in combination with channel state information, and the optimal bandwidth allocation proportion is solved by Karush-Kuhn-Tucker condition;Scheduled client uploads model update using allocated bandwidth, and server weighted aggregation;The present application solves the problem that client scheduling and bandwidth allocation are decoupled in existing federal learning, ignore data heterogeneity and channel difference, realizes the collaborative optimization of scheduling and bandwidth allocation, significantly improves the global model convergence accuracy and robustness under non-independent and identically distributed data and bandwidth limited environment.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication and distributed machine learning technology, and in particular to a heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method. Background Technology

[0002] With the rapid development of Internet of Things (IoT) devices and edge computing technology, massive amounts of data are continuously generated on the smart terminal side. Traditional centralized machine learning paradigms require the raw data to be aggregated to the cloud for training, which not only brings high communication overhead but also causes serious privacy leakage risks. Federated learning, as an emerging distributed learning paradigm, allows edge devices to train models locally and collaboratively optimize the global model by exchanging model parameters rather than raw data, effectively breaking down "data silos" while protecting user privacy.

[0003] However, deploying federated learning in large-scale wireless edge networks faces two core challenges: Statistical heterogeneity means that local datasets on different devices often exhibit highly non-independent and identically distributed (Non-IID) characteristics, which can lead to serious conflicts between model update directions, slowing down global convergence and even causing divergence. Communication heterogeneity arises because wireless bandwidth resources are inherently limited, and physical channel conditions (fading, multipath effects, etc.) vary significantly between different devices. If all eligible devices are allowed to upload models simultaneously, network congestion and excessive transmission delays will inevitably occur.

[0004] Existing client scheduling strategies have significant shortcomings. The FedAvg algorithm uses a random selection mechanism, which is prone to triggering severe model oscillations in Non-IID scenarios. Utility-based strategies, such as Power-of-Choice, rely solely on local loss values ​​to select high-loss clients, making them highly sensitive to noisy data. Importance-Sampling relies solely on gradient norms and ignores update direction, which carries the risk of converging to local optima. Clustering-based strategies, such as Cluster-Sampling, improve representativeness but rely on a single metric and do not consider the underlying communication environment.

[0005] More importantly, most existing scheduling algorithms assume that communication resources are uniform or infinite, completely decoupling "device selection" from "bandwidth allocation". This causes devices with high-value data but poor channel conditions to suffer a lot of transmission delays and eventually be dropped due to timeout, which seriously hinders global convergence in Non-IID environments. Summary of the Invention

[0006] In view of this, the present invention provides a heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method, which solves the technical problems of decoupling client scheduling and bandwidth allocation, discarding high-value clients due to poor channel conditions, and poor global model convergence performance in the prior art, and realizes collaborative optimization of client scheduling and bandwidth allocation.

[0007] To achieve the above objectives, this invention provides a heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method, comprising the following steps: S1. Before each round of communication begins, the parameter server determines the candidate client set from the currently available edge clients and broadcasts the current global model to all candidate clients. Each candidate client performs local model training based on its local private dataset and obtains the corresponding local model update information. S2. Each edge client sends lightweight metadata for edge client evaluation to the parameter server. The lightweight metadata includes local data scale information, local data distribution information, and local model update feature information. S3. The parameter server calculates the comprehensive utility value corresponding to each edge client based on the metadata. S4. The parameter server performs positive truncation and normalization on the comprehensive utility value of each candidate client to obtain the selection probability of each candidate client, and performs random sampling without replacement on the candidate clients according to the selection probability to select a preset number of clients to form the participating client set for the current communication round. S5. Before bandwidth optimization, the parameter server obtains the spectral efficiency of the participating clients and uses the spectral efficiency as a fixed parameter that does not change with the bandwidth allocation ratio during this round of bandwidth optimization. Based on the participating clients, the comprehensive utility value after positive value truncation, and the fixed spectral efficiency, a bandwidth resource optimization model is established to determine the corresponding bandwidth resource allocation ratio. S6. The parameter server receives the local model updates uploaded by the scheduled edge client using the allocated bandwidth resources, performs weighted aggregation, and generates the next round of global model.

[0008] Preferably, the comprehensive utility value includes data size utility, data diversity utility, and dynamic gradient projection utility, expressed as: , in, Indicates the first Wheel Client The overall utility value, Indicates the utility of data diversity. Indicates the utility of data scale. This represents the effect of dynamic gradient projection. , , All represent non-negative weight coefficients and satisfy:

[0009] The parameter server truncates the overall utility value to positive values ​​and calculates the client's selection probability, expressed as follows: , , in, Indicates a preset positive number. This represents the overall utility value of candidate client k in round t. This represents the corresponding choice probability. Denotes the set of candidate clients in round t; The default quantity expression is: , in, Indicates the preset scheduling rate; The non-replacement random sampling renormalizes the selection probability of the remaining candidate clients after each sampling until a client is selected. One client.

[0010] Preferably, the data scale utility The expression is: , in, This indicates the number of local samples on the client side. This represents the maximum number of samples across all clients, and the utility of the data scale. Normalize to a preset range; The utility of data diversity The Shannon entropy measure is used to determine the class distribution balance of the local dataset, expressed as: , in, This represents the total number of categories in the local dataset. Indicates client Belongs to the category The sample proportion To represent small positive numbers, thus avoiding numerical instability; The dynamic gradient projection effect The expression is: , in, Indicates client In the Local gradient update in the round, Indicates the first The global update direction of the wheel, Describing the vector norm, This represents the cosine similarity.

[0011] Preferably, the bandwidth resource optimization model aims to minimize the utility-weighted communication delay, and its expression is: , , in, Indicates assignment to the client bandwidth ratio, This represents the overall utility value after positive value truncation. Indicates the total system bandwidth. Indicates the size of the model update data. Indicates client The spectral efficiency is obtained before the bandwidth optimization of the current communication round and remains constant during this round of solution. Indicates the first A set of clients participating in model training and parameter uploading in rounds; The constraint expression is: ,

[0012] Preferably, the spectral efficiency The expression is: , , in, This represents the equivalent signal-to-noise ratio measured or predicted under a fixed reference configuration. In the current bandwidth optimization process, without following The changes, the client's uplink transmission rate, and the model transmission delay expression are as follows: , , in, Indicates the uplink speed. This indicates the model transmission delay.

[0013] Preferably, the bandwidth resource allocation ratio is obtained through optimization based on the Karush-Kuhn-Tucker conditions, and the expression for the optimal bandwidth allocation ratio is:

[0014] Preferably, the bandwidth resource optimization model further includes a deadline constraint, whereby the total completion time of the client meets the following conditions. At that time, the client can participate in this round of aggregation; in, Indicates local computation latency. Indicates transmission delay. Indicates the preset communication wheel The deadline.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention overcomes the shortcomings of existing single-index scheduling strategies that are prone to getting trapped in local optima or model oscillations in scenarios with non-independent and identically distributed data by constructing a three-dimensional comprehensive utility evaluation system that integrates data scale, data distribution balance, and dynamic gradient projection direction. Based on this, the invention incorporates client scheduling and bandwidth resource allocation into a unified optimization framework. In the scheduling phase, clients with high comprehensive utility are prioritized, while in the bandwidth allocation phase, the goal is to minimize utility-weighted communication latency and obtain a closed-form optimal solution using the Karush-Kuhn-Tucker condition. This achieves a synergistic trade-off between training value and channel quality while avoiding the additional overhead of complex iterative calculations, making it suitable for real-time edge deployment scenarios. Simultaneously, the invention introduces a communication deadline constraint to ensure that all selected clients can complete the upload within the time limit, enhancing engineering feasibility in dynamic wireless environments. The global model test accuracy of the proposed strategy is significantly better than that of uniform allocation and channel quality-first schemes. Even under extreme conditions where nearly 30% of updates are discarded due to timeouts, the accuracy remains at the optimal level, fully verifying its substantial improvement in global convergence performance and robustness under the dual challenges of heterogeneous data and bandwidth constraints. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0018] Example 1: This embodiment provides a heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method, such as... Figure 1 As shown, it includes the following steps: S1. Before each round of communication begins, the parameter server determines the candidate client set from the currently available edge clients and broadcasts the current global model to all candidate clients. Each candidate client performs local model training based on its local private dataset and obtains the corresponding local model update information. S2. Each edge client sends lightweight metadata for edge client evaluation to the parameter server. The lightweight metadata includes local data scale information, local data distribution information, and local model update feature information. S3. The parameter server calculates the comprehensive utility value corresponding to each edge client based on the metadata. The overall utility value includes data size utility, data diversity utility, and dynamic gradient projection utility, expressed as: , in, Indicates the first Wheel Client The overall utility value, Indicates the utility of data diversity. Indicates the utility of data scale. This represents the effect of dynamic gradient projection. , , All represent non-negative weight coefficients and satisfy:

[0019] The parameter server truncates the overall utility value to positive values ​​and calculates the client's selection probability, expressed as: , , in, Indicates a preset positive number. This represents the overall utility value of candidate client k in round t. This represents the corresponding choice probability. Denotes the set of candidate clients in round t; The default quantity expression is: , in, Indicates the preset scheduling rate; The non-replacement random sampling renormalizes the selection probability of the remaining candidate clients after each sampling until a client is selected. One client; Data scale utility The expression is: , in, This indicates the number of local samples on the client side. This represents the maximum number of samples across all clients, and the utility of the data scale. Normalize to a preset range; The utility of data diversity The Shannon entropy measure is used to determine the class distribution balance of the local dataset, expressed as: , in, This represents the total number of categories in the local dataset. Indicates client Belongs to the category The sample proportion Representing small positive numbers avoids numerical instability and enhances data diversity. The larger the value, the more likely it is to be a client The more balanced the category distribution, the stronger its representativeness to the global model; The dynamic gradient projection effect The expression is: , in, Indicates client In the Local gradient update in the round, Indicates the first The global update direction of the wheel, Describing the vector norm, Represents cosine similarity; dynamic gradient projection utility Simultaneously considering the novelty of the magnitude and direction of local updates, clients that provide large and complementary directional updates can achieve more efficient utilization.

[0020] S4. The parameter server performs positive truncation and normalization on the comprehensive utility value of each candidate client to obtain the selection probability of each candidate client, and performs random sampling without replacement on the candidate clients according to the selection probability to select a preset number of clients to form the participating client set for the current communication round. When deploying wirelessly according to this method, candidate clients first report lightweight metadata. After the parameter server completes probability sampling and bandwidth allocation, only selected clients upload complete model updates. In the single-process simulation framework used to verify this method, the local model object of the candidate client can be materialized in memory by the framework first. The parameter server calculates the dynamic utility corresponding to the metadata based on this and then performs the same probability sampling. This memory materialization does not indicate that the candidate client uploads the complete model on the real wireless link before selection, nor is it used to statistically analyze or prove the control signaling communication overhead. Both implementations use the same formulas for overall utility, selection probability, fixed spectral efficiency within a round, and bandwidth allocation.

[0021] S5. Before bandwidth optimization, the parameter server obtains the spectral efficiency of the participating clients and uses the spectral efficiency as a fixed parameter that does not change with the bandwidth allocation ratio during this round of bandwidth optimization. Based on the participating clients, the comprehensive utility value after positive value truncation, and the fixed spectral efficiency, a bandwidth resource optimization model is established to determine the corresponding bandwidth resource allocation ratio. The bandwidth resource optimization model aims to minimize the utility-weighted communication delay, and its expression is: , , in, Indicates assignment to the client bandwidth ratio, This represents the overall utility value after positive value truncation. This represents the total system bandwidth, measured in Hz. This indicates the size of the model update data, in bits. Indicates client The spectral efficiency, measured in bits / s / Hz, is obtained before the current communication round's bandwidth optimization and remains constant throughout this round of solution. Indicates the first A set of clients participating in model training and parameter uploading in rounds; The optimization variables of the bandwidth resource optimization model are only the bandwidth allocation ratio of each participating client; the comprehensive utility value after positive value truncation, spectral efficiency, total system bandwidth and model update data size are all determined before the solution begins and remain unchanged during this round of solution. The constraint expression is: ,

[0022] In this embodiment, the reference configuration is a measurement configuration predetermined before the start of bandwidth optimization and remaining unchanged during this round of optimization. Its configuration parameters do not include the bandwidth allocation ratio to be determined in this round, and the equivalent signal-to-noise ratio obtained from the reference configuration does not change with the bandwidth allocation ratio. The parameter server obtains the equivalent signal-to-noise ratio of each participating client and determines the spectral efficiency before the bandwidth optimization of the current communication round. During this round of bandwidth ratio calculation, this spectral efficiency remains fixed as an exogenous parameter. The expression is: , , in, This represents the equivalent signal-to-noise ratio measured or predicted under a fixed reference configuration. In the current bandwidth optimization process, without following The changes, the client's uplink transmission rate, and the model transmission delay expression are as follows: , , in, This indicates the uplink speed, measured in bits per second. This represents the model transmission delay, measured in seconds (s).

[0023] To correspond to the dynamic channel simulation in the actual code, this embodiment first generates a basic spectral efficiency for each client, and then adds an independent Gaussian perturbation before each round of optimization and truncates it to a preset range: , ,

[0024] The random disturbance only reflects channel changes between communication rounds; after bandwidth optimization in the current round, Remains unchanged. This processing is consistent with... The closed-form derivation for solving the bandwidth ratio is consistent; To solve the bandwidth resource optimization model, let The solution is obtained by using the Lagrange multiplier method combined with the Karush-Kuhn-Tucker conditions. For any participating client, the second derivative of the objective function with respect to the corresponding bandwidth allocation ratio is: , Since both the coefficients and the bandwidth allocation ratio are strictly positive, the above equation is always greater than 0. Therefore, the objective function is strictly convex in the feasible region. The linear equality constraints that allocate all bandwidth constitute a convex set. Therefore, the stationary point determined by the Karush-Kuhn-Tucker conditions is the unique global optimal solution. The expression for the Lagrange constructor is: , in, Representing Lagrange multipliers, since Fixed during this round of solution, i.e. According to the Karush-Kuhn-Tucker conditions Taking the partial derivative and setting it to zero, the expression is: , Solve have to: , Substituting the constraint that all bandwidth is allocated, we get: , Thus, the square root of the Lagrange multiplier can be obtained: , Substitute it back into the stationary point expression and substitute it into... Definition; by canceling out the common constant S / B, the optimal bandwidth allocation ratio is obtained: , The bandwidth resource optimization model also includes a deadline constraint, when the client's total completion time meets the following conditions: At that time, the client can participate in this round of aggregation; in, Indicates local computation latency. Indicates transmission delay. Indicates the preset communication wheel The deadline.

[0025] S6. The parameter server receives the local model updates uploaded by the scheduled edge client using the allocated bandwidth resources, performs weighted aggregation, and generates the next round of global model.

[0026] Example 2: This embodiment uses a lightweight CNN model on the public dataset FMNIST. The total system bandwidth is set to 10 MHz, the model update data size is S, and three communication deadline thresholds of 2500ms, 2000ms, and 1500ms are set. The bandwidth allocation strategy based on convex optimization (referred to as "this invention" in the table) of this invention is compared with the uniform allocation strategy (Equal) and the channel quality-first linear allocation strategy (Linear). The scheduling success rate refers to the proportion of clients that successfully upload model updates within the deadline, and the test accuracy is the highest accuracy of the global model on the test set. The experimental results are shown in Table 1: Table 1

[0027] As shown in Table 1, under the three time thresholds, the linear allocation strategy prioritizing channel quality consistently maintains a 100% scheduling success rate, but its test accuracy is the lowest (approximately 82.2%) across all thresholds. This is because the strategy only favors clients with good channel conditions, completely ignoring the training value of local data, leading to a large number of low-value or even redundant updates participating in aggregation, ultimately resulting in a local optimum. The uniform allocation strategy (Equal) has a scheduling success rate of 80.17% and an accuracy of 83.43% under the relatively lenient 2500ms threshold; however, when the threshold is tightened to 1500ms, its scheduling success rate drops to 74.51%, and its accuracy also decreases to 82.82%, indicating that static uniform allocation is difficult to adapt to dynamic wireless environments. In comparison, the convex optimization bandwidth allocation strategy proposed in this invention achieved the highest test accuracy across all time thresholds. Particularly at the 2000 ms threshold, the accuracy reached 84.79%, significantly outperforming Linear's 82.24% and Equal's 83.45%. Even at the extremely stringent 1500 ms threshold, the scheduling success rate dropped to 70.71% (i.e., nearly 30% of updates were discarded due to timeouts), but its test accuracy still reached 83.84%, higher than Linear's 82.25% and Equal's 82.82%. This fully demonstrates that by jointly considering client utility and wireless channel quality, this invention prioritizes the allocation of limited bandwidth resources to high-value updates, effectively ensuring the convergence performance and robustness of the global model in Non-IID scenarios.

[0028] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method, characterized in that, Includes the following steps: S1. Before each round of communication begins, the parameter server determines the candidate client set from the currently available edge clients and broadcasts the current global model to all candidate clients. Each candidate client performs local model training based on its local private dataset and obtains the corresponding local model update information. S2. Each edge client sends lightweight metadata for edge client evaluation to the parameter server. The lightweight metadata includes local data scale information, local data distribution information, and local model update feature information. S3. The parameter server calculates the comprehensive utility value corresponding to each edge client based on the metadata. S4. The parameter server performs positive truncation and normalization on the comprehensive utility value of each candidate client to obtain the selection probability of each candidate client, and performs random sampling without replacement on the candidate clients according to the selection probability to select a preset number of clients to form the participating client set for the current communication round. S5. Before bandwidth optimization, the parameter server obtains the spectral efficiency of the participating clients and uses the spectral efficiency as a fixed parameter that does not change with the bandwidth allocation ratio during this round of bandwidth optimization. Based on the participating clients, the comprehensive utility value after positive value truncation, and the fixed spectral efficiency, a bandwidth resource optimization model is established to determine the corresponding bandwidth resource allocation ratio. S6. The parameter server receives the local model updates uploaded by the scheduled edge client using the allocated bandwidth resources, performs weighted aggregation, and generates the next round of global model.

2. The heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method according to claim 1, characterized in that, The comprehensive utility value includes data size utility, data diversity utility, and dynamic gradient projection utility, expressed as: ， in, Indicates the first Wheel Client The overall utility value, Indicates the utility of data diversity. Indicates the utility of data scale. This represents the effect of dynamic gradient projection. , , All represent non-negative weight coefficients and satisfy: ， The parameter server truncates the overall utility value to positive values ​​and calculates the client's selection probability, expressed as follows: ， ， in, Indicates a preset positive number. This represents the overall utility value of candidate client k in round t. This represents the corresponding choice probability. Denotes the set of candidate clients in round t; The default quantity expression is: ， in, Indicates the preset scheduling rate; The non-replacement random sampling renormalizes the selection probability of the remaining candidate clients after each sampling until a client is selected. One client.

3. The heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method according to claim 2, characterized in that, The data scale utility The expression is: ， in, This indicates the number of local samples on the client side. This represents the maximum number of samples across all clients, and the utility of the data scale. Normalize to a preset range; The utility of data diversity The Shannon entropy measure is used to assess the class distribution balance of the local dataset, expressed as: ， in, This represents the total number of categories in the local dataset. Indicates client Belongs to the category The sample proportion To represent small positive numbers, thus avoiding numerical instability; The dynamic gradient projection effect The expression is: ， in, Indicates client In the Local gradient updates for each round. Indicates the first The global update direction of the wheel, Describing the vector norm, This represents the cosine similarity.

4. The heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method according to claim 1, characterized in that, The bandwidth resource optimization model aims to minimize the utility-weighted communication delay, and its expression is: ， ， in, Indicates assignment to the client bandwidth ratio, This represents the overall utility value after positive value truncation. Indicates the total system bandwidth. Indicates the size of the model update data. Indicates client The spectral efficiency is obtained before the bandwidth optimization of the current communication round and remains constant during this round of solution. Indicates the first A set of clients participating in model training and parameter uploading in rounds; The constraint expression is: ， 。 5. The heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method according to claim 4, characterized in that, The spectral efficiency The expression is: ， ， in, This represents the equivalent signal-to-noise ratio measured or predicted under a fixed reference configuration. In the current bandwidth optimization process, without following The changes, the client's uplink transmission rate, and the model transmission delay expression are as follows: ， ， in, Indicates the uplink speed. This indicates the model transmission delay.

6. The heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method according to claim 4, characterized in that, The bandwidth resource allocation ratio is obtained through optimization based on the Karush-Kuhn-Tucker conditions, and the expression for the optimal bandwidth allocation ratio is: 。 7. The heterogeneous wireless federated learning utility-aware scheduling and bandwidth allocation method according to claim 1, characterized in that, The bandwidth resource optimization model also includes a deadline constraint, when the client's total completion time meets the following conditions: At that time, the client can participate in this round of aggregation; in, Indicates local computation latency. Indicates transmission delay. Indicates the preset communication wheel The deadline.

Images