A clustered wireless communication method and system

Abstract

The present invention discloses a clustered wireless communication method and system, comprising: step 1, a base station clusters all edge devices participating in the communication into multiple clusters according to the geographical location, and the edge devices in each cluster communicate using a D2D link; step 2, after each edge device updates the decentralized model parameters according to the local data, the edge device exchanges and updates the model parameters with the edge devices connected to it through the D2D link, and uses the parameter averaging algorithm to aggregate the parameters in the cluster to obtain the aggregated parameters; step 3, the base station selects an edge device from all edge devices in each cluster as a cluster head, and the cluster head uploads its aggregated parameters to the base station; step 4, the base station globally aggregates the aggregated parameters uploaded by all cluster heads, and broadcasts the obtained global model to all edge devices. The problem of central node communication traffic congestion in the federated learning scenario with large-scale terminal participation is solved, and the model convergence rate is improved.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a clustered wireless communication method and system. Background Art

[0002] In recent years, with the continuous development of artificial intelligence and big data technology, how to use machine learning algorithms to effectively mine and apply information on massive data has received a lot of attention and research. Typical machine learning algorithms usually require the training data to be transmitted to a base station equipped with a base station, and then the machine learning model is trained centrally. However, as more and more edge devices, such as mobile phones, smart wearable devices, optical sensors, and surveillance cameras are connected to the Internet, the data used to train the model may be collected by various devices. Due to data privacy and security issues in transmission, it is difficult to upload all the collected training data to the base station through wireless communication networks. Therefore, how to securely access data on heterogeneous devices to effectively train models has become a research problem that needs to be solved urgently.

[0003] Research on the Federated Learning (FL) framework in distributed learning theory has become a rapidly developing topic in the scientific and technological community. It provides a new way of model learning, using a large amount of data distributed on a large number of devices to collaboratively train high-quality global models. It distributes model training to several terminals. The terminals do not need to send the original data to the base station. Instead, they upload the model parameters or gradient information after training the model with local data. In this way, data privacy and communication efficiency can be maintained on a highly distributed set of devices.

[0004] With the continuous development of mobile communications and the rapid growth of mobile smart devices, the computing power of terminal devices has been greatly improved, which can be used for model training. Therefore, combining federated learning with wireless communication networks can make full use of the computing resources of terminal nodes and perform more effective machine learning while ensuring user privacy. However, due to limited bandwidth resources and unstable wireless channels, the transmission of a large number of model parameters will bring huge communication overhead and cause large communication delays.

[0005] In traditional federated learning scenarios, terminal devices periodically upload their local gradients or local models to the base station, which aggregates them to obtain a high-quality global model. This framework relies on the base station to collect model parameters or gradients. However, considering practical applications, in scenarios where large-scale terminals jointly participate in federated learning training, communication resources are a key factor affecting the convergence rate due to the large amount of model information exchanged between edge devices and base stations. The traditional federated learning architecture will cause communication traffic congestion problems at the central node. When the available network bandwidth in the system is small, the convergence rate will drop significantly. Therefore, how to solve the communication traffic congestion of the central node of federated learning in a reasonable way is an issue that urgently needs to be studied. Summary of the invention

[0006] In view of the above, the purpose of the present invention is to provide a clustered wireless communication method and system, which can solve the problem of base station communication traffic congestion in large-scale federated learning, reduce the communication delay in the training process, and improve the convergence rate.

[0007] To achieve the above-mentioned purpose of the invention, the embodiment provides the following technical solutions:

[0008] In a first aspect, an embodiment provides a clustered wireless communication method, wherein a system implementing the method includes a base station and a plurality of edge devices communicating with the base station, including the following steps:

[0009] Step 1: The base station clusters all edge devices participating in the communication into multiple clusters based on their geographical locations, and the edge devices in each cluster communicate using D2D links.

[0010] Step 2: After each edge device updates the model parameters by gradient descent based on local data, the edge device exchanges the updated model parameters with the edge devices connected to it through the D2D link, and uses the parameter averaging algorithm to aggregate the parameters within the cluster to obtain the aggregated parameters.

[0011] Step 3: The base station selects an edge device from all edge devices in each cluster as a cluster head, and the cluster head uploads its aggregation parameters to the base station;

[0012] In step 4, the base station globally aggregates the aggregation parameters uploaded by all cluster heads and broadcasts the obtained global model to all edge devices.

[0013] In step 1 of one embodiment, a device connectivity graph is constructed for each cluster, wherein edge devices are used as nodes, and edges are constructed between nodes based on D2D links between edge devices to form a device connectivity graph. The adjacency matrix corresponding to the device connectivity graph represents the connectivity of D2D links between nodes, wherein an element value in the adjacency matrix is ​​1, indicating that the D2D link between two edge devices is connected, and an element value is 0, indicating that the D2D link between the two edge devices is not connected.

[0014] In step 2 of one embodiment, parameter averaging algorithm is used to perform parameter aggregation within a cluster, including:

[0015]

[0016] Where i and j are the indexes of edge devices, ω i Represents the aggregation parameters, represents the model parameters before updating, n k represents the number of edge devices in cluster k, a j,i Indicates the connectivity of the D2D link between edge device i and edge device j:

[0017]

[0018] In step 2 of an embodiment, all D2D communications within the cluster share bandwidth resources B (1) , the optimal bandwidth resource allocation to minimize the communication time required for cluster aggregation is:

[0019]

[0020] Where k is the cluster index, l is the number of iterations, i,j represents the index of the edge device, and M is the number of quantized bits of the model parameters. is the D2D link e transmitted from edge device i to edge device j i,j The allocated bandwidth, P i k,l is the transmit power of terminal device i, For D2D link i,j The channel gain, N0 is the variance of the additive white Gaussian noise, E k It is the set of edges in the device connectivity graph corresponding to cluster k.

[0021] In step 3 of one embodiment, selection is made based on the connectivity of nodes in the device connectivity graph corresponding to the cluster, that is, the edge device corresponding to the node with the largest connectivity in each cluster is selected as the cluster head, including:

[0022]

[0023] Where D(v) refers to the degree of node v in the device connectivity graph, that is, the number of edges connected to node v, V k Represents the set of nodes in the device connectivity graph corresponding to cluster k.

[0024] In step 3 of an embodiment, selecting a cluster head according to the uplink channel state between the edge device and the base station and the transmit power of the edge device includes:

[0025] First, the data transmission rate R between edge device i and the base station in cluster k k,i It is expressed as follows:

[0026]

[0027] Then, the communication time T required for edge device i in cluster k to transmit model parameters to the base station is i (2) It is expressed as follows:

[0028]

[0029] Based on this, the cluster head s corresponding to cluster k k The selection method is as follows:

[0030]

[0031] Among them, B k represents the bandwidth allocated for communication between edge devices in cluster k and the base station, P k,i is the transmit power of edge device i, h k,i is the uplink channel gain of edge device i, M is the amount of model parameter data that needs to be transmitted, and n k Represents the total number of edge devices in cluster k.

[0032] In step 3 of an embodiment, the bandwidth resource allocated for communication between the cluster head and the base station is B (2) In order to minimize the communication time required for global aggregation, bandwidth resource B (2) The optimal allocation is as follows:

[0033]

[0034] Among them, k represents the index of the cluster, B k * indicates cluster head s k The allocated bandwidth, P k is the cluster head k Transmitting power, h k is the cluster head k is the uplink channel gain, N0 is the variance of additive white Gaussian noise, and M is the number of quantization bits of the model parameters.

[0035] In a second aspect, an embodiment provides a clustered wireless communication system, including a base station, and a plurality of edge devices communicating with the base station.

[0036] The base station clusters all edge devices participating in the communication into multiple clusters according to their geographical locations, and the edge devices in each cluster communicate using D2D links;

[0037] After each edge device updates the model parameters by gradient descent according to local data, the edge device exchanges the updated model parameters with the edge device connected to it through the D2D link, and uses the parameter averaging algorithm to aggregate the parameters within the cluster to obtain the aggregated parameters;

[0038] The base station selects an edge device from all edge devices in each cluster as a cluster head, and the cluster head uploads its aggregation parameters to the base station;

[0039] The base station globally aggregates the aggregation parameters uploaded by all cluster heads, and broadcasts the obtained global model to all edge devices.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] According to the geographical location of the edge devices, a clustering algorithm is used to divide the edge devices into several clusters, and decentralized training is performed in parallel in each cluster, which effectively utilizes the wireless D2D network in the area and solves the problems of limited communication resources and base station traffic congestion. There is only one cluster head in each cluster that transmits model parameters to the central server, which effectively reduces the amount of data communicated between the edge devices and the base station, reduces the communication time, and improves the convergence rate of the model to reach convergence. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 is a schematic structural diagram of a clustered wireless communication system provided in an embodiment;

[0044] Figure 2 is a flow chart of a clustered wireless communication method provided by an embodiment;

[0045] Figure 3 is a schematic diagram of wireless communication interaction provided by an embodiment;

[0046] Figure 4 This is a schematic diagram of wireless communication interaction provided by another embodiment. DETAILED DESCRIPTION

[0047] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0048] Figure 1 It is a structural diagram of a clustered wireless communication system provided in an embodiment. The wireless communication system provided in the embodiment includes a base station equipped with a central server, and multiple edge devices communicating with the base station as terminals. Among them, each edge device maintains local data for model training, and the local data and the base station jointly complete the training of the model. Federated learning with the participation of large-scale edge devices can be realized by using this clustered wireless communication system. Specifically, after the edge devices are clustered according to their geographical locations, each cluster independently performs decentralized training and aggregates model parameters based on the D2D network. A cluster head is selected from each cluster according to the cluster head selection strategy, and its aggregation parameters are uploaded to the base station for global model aggregation. The problem of traffic congestion on the central server is solved. The amount of data exchanged between the terminal device and the base station is effectively reduced, the communication delay during the training process is reduced, and the model convergence rate is improved.

[0049] Figure 2 FIG. 1 is a flow chart of a clustered wireless communication method provided by an embodiment. Figure 2 As shown, the clustered wireless communication method provided in the embodiment is applied Figure 1 The system shown comprises the following steps:

[0050] In step 201, the base station groups all edge devices participating in the communication into multiple clusters according to their geographical locations, and the edge devices in each cluster communicate using D2D links.

[0051] In the embodiment, the base station uses a clustering algorithm to cluster all edge devices into K clusters according to their address locations. The edge devices in the same cluster perform decentralized machine learning training, and the communication between edge devices is based on a wireless D2D network.

[0052] Due to factors such as path loss and noise, the D2D links between some devices are blocked. Therefore, a device connectivity graph is constructed for each cluster to characterize the D2D link communication of edge devices within the cluster. The connectivity structure between devices within the cluster remains unchanged during the entire training process. Specifically, edge devices are used as nodes, and edges are constructed between nodes based on the D2D links between edge devices to form a device connectivity graph G. k (V k ,E k ), where V k Represents a collection of nodes. Represents the edge set. The matrix corresponding to the device connectivity graph represents the connectivity of the D2D links between nodes, where the element value in the matrix is ​​1, indicating that the D2D link between the two edge devices is connected, and the element value is 0, indicating that the D2D link between the two edge devices is not connected.

[0053] In step 202, after each edge device updates the model parameters by gradient descent according to local data, the edge device exchanges the updated model parameters with the edge devices connected to it through the D2D link, and uses the parameter averaging algorithm to aggregate the parameters within the cluster to obtain the aggregated parameters.

[0054] In the embodiment, the edge device i in cluster k uses the stochastic gradient descent algorithm to update the model parameters according to the local data set D i A randomly sampled data sample ξ t,l The model parameters are updated by gradient descent, and the model parameters obtained in the lth iteration of the tth round of training are It is expressed as:

[0055]

[0056] in, represents the model parameters obtained in the l-1th iteration, η t,l is the learning rate of the lth iteration of the tth round of training, ▽ is the gradient operator, and L is the loss function.

[0057] In the embodiment, edge device i in cluster k exchanges model parameters with edge devices connected to it based on D2D network communication, aggregates the model parameters in the cluster, and the aggregated parameters of edge device i after averaging the model parameters are used as the starting point of the l+1th iteration. It is expressed as:

[0058]

[0059] Among them, n k represents the number of edge devices in cluster k, a j,i Indicates the connectivity of the D2D link between edge device i and edge device j:

[0060]

[0061] In the embodiment, the available bandwidth resource of the wireless communication system is B. For the decentralized training phase within the cluster, all D2D communications within the cluster share the bandwidth resource B. (1) The bandwidth resource used by the cluster head and the base station for communication is B (2) .

[0062] B=B (1) +B (2)

[0063] Each terminal device transmits the model parameters to the edge devices connected to it in the device connectivity graph. Due to the long geographical distance between clusters, the interference between the communications of edge devices in different clusters can be regarded as noise.

[0064] For cluster k, frequency division multiple access (FDMA) is used to allocate B (1) It is divided into several sub-channels, each of which is assigned to a directed edge (i→j) in the device connectivity graph, denoted as e i,j Considering the lth D2D communication within cluster k, the D2D link e transmitted from edge device i to edge device j is i,j The allocated bandwidth is The relationship is as follows:

[0065]

[0066] The transmission rate of communication when edge device i transmits model parameters to edge device j It is expressed as follows:

[0067]

[0068] Among them, P i k,l is the transmit power of terminal device i, For D2D link i,j is the channel gain, and N0 is the variance of the additive white Gaussian noise.

[0069] Assuming that the amount of data of the model parameters to be transmitted each time is M bits, the maximum time required for the lth D2D communication between edge device i and edge device j is It is expressed as:

[0070]

[0071] The time required for cluster k to perform L rounds of decentralized training is accumulated. It is expressed as follows:

[0072]

[0073] The communication time required for all K clusters to complete L rounds of intra-cluster decentralized training is T (1) It is expressed as follows:

[0074]

[0075] In order to minimize the communication time T required for cluster aggregation (1) , for bandwidth resource B (1) The optimal allocation is as follows:

[0076]

[0077] Step 203 , the base station selects an edge device from all edge devices in each cluster as a cluster head, and the cluster head uploads its aggregation parameters to the base station.

[0078] In the embodiment, after performing L model parameter updates and cluster model parameter aggregation in K clusters using the above step 2, an edge device is selected as a cluster head in each cluster according to the cluster head selection strategy. The set of cluster head devices is represented as S = {s1, s2, ..., s K}.

[0079] In one possible implementation, the cluster head selection strategy adopted is to select based on the connectivity of the nodes in the device connectivity graph within the cluster. The connectivity D(v) of node v in the device connectivity graph refers to the number of edges connected to node v. The larger D(v) is, the more other nodes are connected to the node, that is, the greater the connectivity of the node represented by the edge device, and it can exchange model parameters with more devices when performing parameter aggregation within the cluster. After parameter averaging, a more accurate aggregation parameter representing the model after local training of the devices in the cluster can be obtained. Therefore, the edge device with the largest connectivity in each cluster is selected as the cluster head, and the aggregation parameters are uploaded to the base station. Cluster head s in cluster k k The selection method is as follows:

[0080]

[0081] In another possible real-time mode, the cluster head selection strategy is to select according to the uplink channel status between the edge device and the base station and the transmission power of the edge device. Specifically, considering the communication delay of the system, the bandwidth resource allocated for communication between the device and the server is B (2) , the data transmission rate R between edge device i and base station in cluster k k,i It is expressed as follows:

[0082]

[0083] Among them, B k represents the bandwidth allocated for communication between edge devices in cluster k and the base station, P k,i is the transmit power of edge device i, h k,i is the uplink channel gain of edge device i, then the communication time between edge device i and base station in cluster k is T i (2) It is expressed as follows:

[0084]

[0085] Wherein, M is the amount of model parameter data that needs to be transmitted.

[0086] Communication time T of edge device i i (2)The shorter it is, the better the communication quality of the uplink channel is, which can reduce the training time required for the overall federated learning to converge and improve the overall convergence rate of the model. i (2) It depends on the transmission power of the device at that moment and its uplink channel gain. Therefore, the cluster head selection strategy selects the edge device with the best channel status in each cluster as the cluster head and sends the aggregation parameters to the base station. k The selection method is as follows:

[0087]

[0088] In the embodiment, K cluster heads upload their intra-cluster aggregation parameters to the base station, and the bandwidth resources allocated for communication between the cluster heads and the base station are B (2) .

[0089] Orthogonal Frequency Division Multiple Access (OFDMA) is used to (2) It is divided into K sub-channels. Cluster head s k The allocated bandwidth is B k , then the relationship is as follows:

[0090]

[0091] Cluster heads k Data transmission rate between the base station and the k It is expressed as follows:

[0092]

[0093] Among them, P k is the cluster head k Transmitting power, h k is the cluster head k is the uplink channel gain, and N0 is the variance of the additive white Gaussian noise.

[0094] The communication time required for a global aggregation is T (2) It is expressed as follows:

[0095]

[0096] In order to minimize the communication time T required for global aggregation (2) , for bandwidth resource B (2) The optimal allocation is as follows:

[0097]

[0098] In step 204, the base station globally aggregates the aggregation parameters uploaded by all cluster heads, and broadcasts the obtained global model to all edge devices.

[0099] In the embodiment, the K cluster heads respectively use their trained aggregation parameters Uploaded to the base station, the base station averages the K aggregation parameters to obtain the global model ω t+1 Expressed as:

[0100]

[0101] The base station broadcasts the global model ω to all N devices t+1 .

[0102] 205 , repeat the above steps 201 to 204 and perform T rounds of training until the global model converges.

[0103] The above-mentioned clustered wireless communication method divides edge devices into different clusters, solving the problems of limited communication resources and traffic congestion of central servers; independent decentralized training is carried out in each cluster, effectively utilizing the wireless D2D network in the area; and the cluster head device selection strategy is used to select the cluster head to upload aggregation parameters, thereby improving the convergence rate.

[0104] Figure 3 FIG. 1 is a schematic diagram of wireless communication interaction provided by an embodiment. Figure 3 As shown, the following steps are included:

[0105] 301. Perform several decentralized trainings on all clusters in parallel, including using local data for local training and then utilizing wireless D2D communication to aggregate parameters within the cluster.

[0106] 302. All edge devices upload their connectivity information to the base station.

[0107] 303. The scheduler on the base station selects the edge device with the largest connectivity in each cluster as the cluster head according to the collected connectivity information.

[0108] 304. Allocate bandwidth for the uplink, and the cluster head of each cluster uploads aggregation parameters to the base station.

[0109] 305. The base station aggregates the global model by parameter averaging.

[0110] 306. The base station transmits the global model to all terminal devices in a broadcast manner.

[0111] Figure 4 FIG. 1 is a schematic diagram of wireless communication interaction provided by another embodiment. Figure 4 As shown, the following steps are included:

[0112] 401. Perform several decentralized trainings on all clusters in parallel, including using local data for local training and then utilizing wireless D2D communication to aggregate parameters within the cluster.

[0113] 402. All edge device terminals upload their own uplink channel state information to the base station.

[0114] 403. The scheduler on the base station selects the edge device with the best channel state in each cluster as the cluster head according to the collected uplink channel state information.

[0115] 404. Allocate bandwidth for the uplink, and the cluster head of each cluster uploads aggregation parameters to the base station.

[0116] 405. The base station aggregates the global model by parameter averaging.

[0117] 406. The base station transmits the global model to all terminal devices in a broadcasting manner.

[0118] In the embodiment, the communication mode between the edge device and the base station is a wireless communication mode, which can be an existing mobile communication network, i.e., LTE (Long-term Evolution) or 5G network, or a WiFi network.

[0119] In the embodiment, the processor capability of the central server included in the base station far exceeds the computing capability of the edge device, and has the ability to independently perform model training. The edge device can be a modern smart phone, tablet computer, notebook, self-driving car and other mobile terminals that can support model training, equipped with a wireless communication system, and can access mainstream wireless communication networks such as mobile communication networks and WiFi.

[0120] The wireless communication method provided in the above embodiment, in a federated learning scenario, divides the devices into several clusters according to their geographical locations through a clustering algorithm, and performs decentralized training in parallel within each cluster, thereby solving the problem of central node communication traffic congestion in federated learning scenarios with the participation of large-scale terminals, effectively reducing the amount of data exchanged between terminal devices and base stations, reducing communication delays during training, and improving the model convergence rate.

[0121] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A clustered wireless communication method, wherein a system implementing the method comprises a base station and a plurality of edge devices communicating with the base station, wherein: The following steps are involved: Step 1: The base station clusters all edge devices participating in the communication into multiple clusters based on their geographical locations, and the edge devices in each cluster communicate using D2D links. Step 2: After each edge device updates the model parameters by gradient descent based on local data, the edge device exchanges the updated model parameters with the edge devices connected to it through the D2D link, and uses the parameter averaging algorithm to aggregate the parameters within the cluster to obtain the aggregated parameters, including: Where i and j are the indexes of edge devices, ω i Represents the aggregation parameters, represents the model parameters before updating, n k represents the number of edge devices in cluster k, a j,i Indicates the connectivity of the D2D link between edge device i and edge device j: Step 3: The base station selects an edge device from all edge devices in each cluster as a cluster head, and the cluster head uploads its aggregation parameters to the base station; In step 4, the base station globally aggregates the aggregation parameters uploaded by all cluster heads and broadcasts the obtained global model to all edge devices.

2. The clustered wireless communication method according to claim 1, characterized in that: In step 1, a device connectivity graph is constructed for each cluster, wherein edge devices are taken as nodes, and edges are constructed between nodes based on the D2D links between edge devices to form a device connectivity graph. The adjacency matrix corresponding to the device connectivity graph represents the connectivity of the D2D links between nodes, wherein an element value in the adjacency matrix is ​​1, indicating that the D2D link between the two edge devices is connected, and an element value is 0, indicating that the D2D link between the two edge devices is not connected.

3. The clustered wireless communication method according to claim 1, characterized in that: In step 2, all D2D communications in the cluster share bandwidth resources B ( 1 ), the optimal bandwidth resource allocation to minimize the communication time required for cluster aggregation is: Where k is the cluster index, l is the number of iterations, i,j represents the index of the edge device, and M is the number of quantized bits of the model parameters. is the D2D link e transmitted from edge device i to edge device j i,j The allocated bandwidth, P i k,l is the transmit power of terminal device i, For D2D link i,j The channel gain, N0 is the variance of the additive white Gaussian noise, E k It is the set of edges in the device connectivity graph corresponding to cluster k.

4. The clustered wireless communication method according to claim 2, characterized in that: In step 3, the selection is made based on the connectivity of the nodes in the device connectivity graph corresponding to the cluster, that is, the edge device corresponding to the node with the largest connectivity in each cluster is selected as the cluster head, including: Where D(v) refers to the degree of node v in the device connectivity graph, that is, the number of edges connected to node v, V k Represents the set of nodes in the device connectivity graph corresponding to cluster k.

5. The clustered wireless communication method according to claim 1, characterized in that: In step 3, a cluster head is selected according to the uplink channel state between the edge device and the base station and the transmit power of the edge device, including: First, the data transmission rate R between edge device i and the base station in cluster k k,i It is expressed as follows: Then, the communication time T required for edge device i in cluster k to transmit model parameters to the base station is i (2) It is expressed as follows: Based on this, the cluster head s corresponding to cluster k k The selection method is as follows: Among them, B k represents the bandwidth allocated for communication between edge devices in cluster k and the base station, P k,i is the transmit power of edge device i, h k,i is the uplink channel gain of edge device i, M is the amount of model parameter data to be transmitted, n k represents the total number of edge devices in cluster k, and N0 is the variance of the additive white Gaussian noise.

6. The clustered wireless communication method according to claim 1, characterized in that: In step 3, the bandwidth resource allocated for communication between the cluster head and the base station is B( 2 ), in order to minimize the communication time required for global aggregation, bandwidth resource B( 2 ) is optimally allocated as follows: Among them, k represents the index of the cluster, B k * indicates cluster head s k The allocated bandwidth, P k is the cluster head k Transmitting power, h k is the cluster head k is the uplink channel gain, N0 is the variance of additive white Gaussian noise, and M is the number of quantization bits of the model parameters.

7. The clustered wireless communication method according to claim 1, characterized in that: In step 1, a clustering algorithm is used to cluster all edge devices into multiple clusters based on their geographical locations; In step 2, all edge devices in each cluster use the stochastic gradient descent algorithm to update the model parameters.

8. A clustered wireless communication system, comprising a base station and a plurality of edge devices communicating with the base station, characterized in that: The base station clusters all edge devices participating in the communication into multiple clusters according to their geographical locations, and the edge devices in each cluster communicate using D2D links; After each edge device updates the model parameters by gradient descent according to local data, the edge device exchanges the updated model parameters with the edge devices connected to it via the D2D link, and uses the parameter averaging algorithm to aggregate the parameters within the cluster to obtain aggregated parameters, including: Where i and j are the indexes of edge devices, ω i Represents the aggregation parameters, represents the model parameters before updating, n k represents the number of edge devices in cluster k, a j,i Indicates the connectivity of the D2D link between edge device i and edge device j: The base station selects an edge device from all edge devices in each cluster as a cluster head, and the cluster head uploads its aggregation parameters to the base station; The base station globally aggregates the aggregation parameters uploaded by all cluster heads, and broadcasts the obtained global model to all edge devices.

Images