Electronic device and method for spectrum management device
By dividing the federated prediction area in a large-scale network and adopting federated learning and K-DBSCAN algorithms, combining the multi-input, multi-output, long and short-term memory neural network model, the communication delay and model complexity problems caused by global federated learning are solved, and the communication delay and model complexity problems are achieved. Efficient prediction of received power of the main system and spectrum resource perception.
Patent Information
- Application Number
- CN202310519954.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-09
- Publication Date
- 2024-11-12
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In large-scale network scenarios, existing spectrum sharing technology leads to high communication delays and model complexity through global federated learning, making it difficult to effectively predict the received power of the main system in different regions.
By dividing the network into multiple federated prediction areas, using federated learning to independently train and aggregate models within each area, reducing model complexity and improving prediction accuracy, using K-DBSCAN algorithm to determine federated prediction areas, combining multiple inputs and multiple outputs Long-term short-term memory neural network model predicts the main system behavior model.
It realizes more efficient power prediction of the main system in different regions, reduces communication costs and model complexity, and improves the accuracy and reliability of spectrum resource perception.
Smart Images

Figure CN118944783A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of wireless communication technology, and in particular to an electronic device and method for a spectrum management device, and more particularly to more effectively predicting the received power of a primary system in different areas for spectrum resource perception. Background Art
[0002] Existing spectrum sharing technologies predict the occupancy of available spectrum resources based on federated learning. In large-scale network scenarios, global federated learning will lead to high communication delays and high model complexity. Summary of the invention
[0003] A brief overview of the present invention is provided below in order to provide a basic understanding of certain aspects of the present invention. It should be understood that this overview is not an exhaustive overview of the present invention. It is not intended to identify key or important parts of the present invention, nor is it intended to limit the scope of the present invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to a more detailed description discussed later.
[0004] According to one aspect of the present disclosure, an electronic device for a spectrum management device is provided, which includes a processing circuit, and the processing circuit is configured to: for at least one federal prediction area among multiple federal prediction areas constructed based on a secondary system, obtain a main system behavior model for predicting the behavior of a main system in using spectrum resources through federated learning.
[0005] In the embodiments according to the present disclosure, the received power of the primary system in different areas can be predicted more effectively for spectrum resource sensing.
[0006] According to one aspect of the present disclosure, an electronic device for a spectrum management device is provided, which includes a processing circuit, and the processing circuit is configured to: for one federal prediction area among multiple federal prediction areas constructed based on a secondary system, assist in obtaining a main system behavior model for predicting the behavior of the main system in using spectrum resources through federated learning.
[0007] In the embodiments according to the present disclosure, it is possible to assist in more effectively predicting the received power of the primary system in different areas for spectrum resource sensing.
[0008] According to one aspect of the present disclosure, a method for a spectrum management device is provided, comprising: for at least one federal prediction area among multiple federal prediction areas constructed based on a secondary system, the spectrum management device obtains a main system behavior model for predicting the behavior of a main system in using spectrum resources through federated learning.
[0009] According to one aspect of the present disclosure, a method for a spectrum management device is provided, including: for one federal prediction area among multiple federal prediction areas constructed based on a secondary system, assisting in obtaining a main system behavior model for predicting the behavior of the main system using spectrum resources through federated learning.
[0010] According to other aspects of the present invention, a computer program code and a computer program product for implementing the above method for a spectrum management device and a computer-readable storage medium having the computer program code for implementing the above method for a spectrum management device recorded thereon are also provided. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] In order to further illustrate the above and other advantages and features of the present invention, the specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. The accompanying drawings together with the following detailed description are included in this specification and form a part of this specification. Elements with the same function and structure are represented by the same reference numerals. It should be understood that these drawings only describe typical examples of the present invention and should not be regarded as limiting the scope of the present invention. In the drawings:
[0012] Figure 1 A functional module block diagram of an electronic device used for a spectrum management device according to an embodiment of the present disclosure is shown;
[0013] FIG. 2( a ) is a diagram showing an example of a system structure according to an embodiment of the present disclosure, and FIG. 2( b ) is a diagram showing another example of a system structure according to an embodiment of the present disclosure;
[0014] Figure 3 is an example showing how to partition a network according to an embodiment of the present disclosure;
[0015] Figure 4 is an example of a pseudo code illustrating a K-DBSCAN algorithm according to an embodiment of the present disclosure;
[0016] Figure 5(a) and 5(b) is an example showing the effect of using the K-DBSCAN algorithm to determine the federated prediction area according to an embodiment of the present disclosure;
[0017] Figure 6 is an example showing a process of constructing a federated prediction region according to an embodiment of the present disclosure;
[0018] Figure 7 is a schematic diagram showing information interaction for constructing a federated learning group according to an embodiment of the present disclosure;
[0019] Figure 8 An example of selecting a federation aggregation point for performing local model aggregation according to an embodiment of the present disclosure is shown;
[0020] Fig. 9 is an example showing the structure of an LSTM network model according to an embodiment of the present disclosure;
[0021] Fig.10 is an example flow chart showing how to obtain a main system behavior model through federated learning based on a federated prediction region according to an embodiment of the present disclosure;
[0022] FIG11(a) shows the actual received power of the main system and the predicted received power of the main system predicted by the prior art, FIG11(b) shows the actual received power of the main system and the predicted received power of the main system predicted by the weighted aggregation of the quantity, and FIG11(c) shows the error of the predicted received power of the main system predicted by the prior art and the error of the predicted received power of the main system predicted by the weighted aggregation of the quantity;
[0023] Fig.12 A functional module block diagram of an electronic device used for a spectrum management device according to another embodiment of the present disclosure is shown;
[0024] Fig.13 A flowchart of a method for a spectrum management device according to an embodiment of the present disclosure is shown;
[0025] Fig.14 A flowchart of a method for a spectrum management apparatus according to another embodiment of the present disclosure is shown;
[0026] Fig.15 is a block diagram showing a first example of a schematic configuration of an eNB or gNB to which the technology of the present disclosure can be applied;
[0027] Fig.16 is a block diagram showing a second example of a schematic configuration of an eNB or gNB to which the technology of the present disclosure can be applied;
[0028] Fig.17 is a block diagram showing an example of a schematic configuration of a smartphone to which the technology of the present disclosure can be applied;
[0029] Fig.18 is a block diagram showing an example of a schematic configuration of a car navigation device to which the technology of the present disclosure can be applied; and
[0030] Fig.19 is a block diagram of an exemplary structure of a general personal computer in which the method and / or apparatus and / or system according to the embodiments of the present invention can be implemented. DETAILED DESCRIPTION
[0031] Exemplary embodiments of the present invention will be described below in conjunction with the accompanying drawings. For the sake of clarity and conciseness, not all features of actual implementations are described in the specification. However, it should be understood that many implementation-specific decisions must be made in the process of developing any such actual implementation in order to achieve the developer's specific goals, such as meeting those constraints related to the system and business, and these constraints may vary from implementation to implementation. In addition, it should be understood that although the development work may be very complex and time-consuming, it is only a routine task for those skilled in the art who benefit from the present disclosure.
[0032] It is also necessary to explain here that, in order to avoid obscuring the present invention due to unnecessary details, only the device structure and / or processing steps closely related to the solution according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.
[0033] Figure 1 The functional module block diagram of an electronic device 100 used for a spectrum management device according to an embodiment of the present disclosure is shown. The electronic device 100 can be used as a spectrum management device.
[0034] like Figure 1 As shown, the electronic device 100 includes: a federal prediction area obtaining unit 101, which can be configured to obtain at least one federal prediction area among multiple federal prediction areas constructed based on a secondary system (also referred to as a secondary user or SU (Secondary User)); and a main system behavior model obtaining unit 103, which can be configured to obtain a main system behavior model (hereinafter sometimes referred to as a global model) for predicting the behavior of a main system (also referred to as a primary user or PU (Primary User)) in using spectrum resources, and the main system behavior model is obtained through federated learning.
[0035] The obtaining federated prediction area unit 101 and the obtaining main system behavior model unit 103 may be implemented by one or more processing circuits, which may be implemented as a chip or a processor, for example. Figure 1 The various functional units in the electronic device shown in the figure are only logical modules divided according to the specific functions implemented by them, and are not used to limit the specific implementation method.
[0036] The electronic device 100 can be used as a network side device in a wireless communication system, and specifically, for example, can be arranged on the base station side or can be communicatively connected to the base station. Here, it should also be pointed out that the electronic device 100 can be implemented at the chip level, or it can also be implemented at the device level. For example, the electronic device 100 can work as a base station itself, and can also include external devices such as a memory, a transceiver (not shown), etc. The memory can be used to store programs and related data information that need to be executed by the electronic device to implement various functions. The transceiver may include one or more communication interfaces to support communication with different devices (for example, servers, base stations, etc.), and the implementation form of the transceiver is not specifically limited here.
[0037] As an example, the base station may be, for example, an eNB or a gNB.
[0038] The wireless communication system according to the present disclosure may be a 5G NR (New Radio) communication system. Further, the wireless communication system according to the present disclosure may include a non-terrestrial network (NTN). Optionally, the wireless communication system according to the present disclosure may also include a terrestrial network (TN). In addition, those skilled in the art may understand that the wireless communication system according to the present disclosure may also be a 4G or 3G communication system.
[0039] The primary system (which includes base stations and user equipment) has priority access to spectrum resources, while the secondary system (which includes base stations and user equipment) does not have legal spectrum access rights. In the spectrum sharing technical requirements, as long as the aggregate interference intensity generated by all secondary systems to the primary system receiver is lower than the interference protection threshold of the primary system, the secondary system can still use the spectrum resources.
[0040] For example, in centralized federated learning, the electronic device 100, i.e., the spectrum management device, can act as a federated learning server and obtain the main system behavior model through federated learning, or can obtain the main system behavior model from a federated learning server (e.g., NWDAF) without participating in federated learning; in addition, for example, in distributed federated learning, the spectrum management device can act as a federated learning client and obtain the local main system behavior model through federated learning.
[0041] In an embodiment according to the present disclosure, based on the secondary system, the network is divided into multiple federated prediction areas. Instead of performing federated learning for the entire network, a corresponding primary system behavior model is obtained through federated learning for each of at least one of the multiple federated prediction areas. That is, by performing federated learning in each of the above-mentioned federated prediction areas, a corresponding primary system behavior model is obtained, which can be applied to spectrum sharing scenarios in the field of wireless communications. The primary system behavior model can be, for example, a deep learning model, which can be used for primary system power prediction.
[0042] In large-scale network scenarios, global federated learning will lead to high communication delays and high model complexity.
[0043] By dividing the federated prediction area, at least one of the following technical benefits can be achieved: it can reduce communication costs, and by reducing the size of the federated prediction area, the scope of local model data analysis in federated learning can be constrained, thereby reducing the complexity of the model; all base stations in the federated prediction area will participate in federated learning, so the local training data of the base station can be more fully utilized, that is, the model accuracy can be improved while reducing overhead. In addition, the network is divided into multiple federated prediction areas, and the spectrum prediction tasks of the network are divided into multiple federated prediction areas. If a single point failure occurs in the center for model aggregation in a single federated prediction area, other federated prediction areas can still complete federated learning. Compared with other spectrum sensing technologies, it can predict the future receiving power of the main system signal in a specific area instead of simply judging the shutdown of the main system. It can more accurately discover the availability of spectrum resources in spatial and temporal dimensions, so as to more effectively predict the receiving power of the main system in different areas for spectrum resource sensing.
[0044] As an example, the spectrum management device, i.e., the electronic device 100, can use the final global model obtained from the federated learning server as the main system behavior model, wherein the federated learning server aggregates local main system behavior models (hereinafter, sometimes referred to as local models) trained by various federated learning clients participating in the federated learning to obtain the final global model.
[0045] In the following, the federated learning client is sometimes referred to as FL-Client or FL Client, and the federated learning server is sometimes referred to as FL-Sever. Federated learning including the federated learning server and the federated learning client is called centralized federated learning.
[0046] As an example, the federated learning server can be implemented by the NWDAF (Network Data Analytics Function) of the 5G core network.
[0047] As an example, the federated learning server sends a federated learning request to other devices in a federated prediction area associated with the server, and receives a federated learning response from other devices that agree to participate in the federated learning.
[0048] As an example, the federated learning response includes the number of geographic units occupied by user devices in the secondary system associated with the federated learning client.
[0049] As an example, the federated learning server: in each round of training of the federated learning, the local main system behavior models trained by each federated learning client are aggregated based on the weights to obtain a global model, and the global model is sent to each federated learning client so that each federated learning client performs the next round of training based on the global model; and until a predetermined condition is met, the federated learning is stopped to obtain a final global model. As an example, the predetermined condition includes that the final global model reaches a predetermined accuracy (for example, the final global model converges) and / or a predetermined number of rounds of training have been performed.
[0050] FIG. 2( a ) shows an example of a system structure according to an embodiment of the present disclosure.
[0051] Figure 2(a) relies on the federated learning framework in the 3GPP TS 23.288 standard. The base stations in the secondary system upload the collected data to the corresponding NWDAF.
[0052] In Figure 2(a), the primary system is represented by PU and the secondary system is represented by SU. SU can sense the signal of PU and upload the collected information to the device or function it accesses. The system consists of an access layer and a data layer. The bottom layer is the access layer, which represents the access relationship between the user equipment and the base station. The upper layer in the system structure in Figure 2(a) is the abstract data layer. The data layer in Figure 2(a) consists of the core network NWDAF and NRF (Network Repository Function). In Figure 2(a), there is one PU and multiple SUs. The base stations in the PU are represented by BS. The base station set of multiple SUs is {BS1, BS2, BS3, BS4, BS5, BS6, BS7, BS8}. For example and not limitation, the base station set is divided into two federated prediction zones FPZ (Federated Prediction Zone): federated prediction zone 1, namely FPZ_1, and federated prediction zone 2, namely FPZ_2, where FPZ_1 = {BS1, BS2, BS3, BS4}, and FPZ_2 = {BS5, BS6, BS7, BS8}.
[0053] As shown in Figure 2(a), for example, FPZ_1 includes NWDAF_1 and NWDAF_2, base station BS1 uploads the collected data Dataset1 to the corresponding NWDAF_1, base station BS2 uploads the collected data Dataset 2 to the corresponding NWDAF_1, base station BS3 uploads the collected data Dataset 3 to the corresponding NWDAF_2, and base station BS4 uploads the collected data Dataset 4 to the corresponding NWDAF_2; NWDAF_1 and NWDAF_2 act as federated learning clients (for example, marked as FPZ_1Client: {1,2} in the figure) and use the collected data to train the local main system behavior model, and NWDAF_0 acts as a federated learning server (for example, marked as FL-Sever: {0} in the figure) to globally aggregate the local models to obtain the final global model (for example, global model 1). For FPZ_2, base station BS5 uploads the collected data Dataset 5 to the corresponding NWDAF_3, base station BS6 uploads the collected data Dataset 6 to the corresponding NWDAF_3, base station BS7 uploads the collected data Dataset 7 to the corresponding NWDAF_4, and base station BS8 uploads the collected data Dataset 8 to the corresponding NWDAF_4. NWDAF_3 and NWDAF_4, as federated learning clients (e.g., marked as FPZ_2Client: {3,4} in the figure), use the collected data to train the local model, and NWDAF_0, as a federated learning server, globally aggregates the local model to obtain the final global model (e.g., global model 2).
[0054] The electronic device 100 may obtain the global model 1 for FPZ_1 from NWDAF_0 as the main system behavior model for FPZ_1 , and may obtain the global model 2 for FPZ_2 from NWDAF_0 as the main system behavior model for FPZ_2 .
[0055] That is, the electronic device 100 implements primary system power prediction through federated learning for at least one area in FPZ_1 and FPZ_2, thereby providing reliable primary system behavior data for spectrum sharing.
[0056] As an example, the spectrum management device, i.e., the electronic device 100, can serve as a federated learning server, and the main system behavior model obtaining unit 103 can be configured to aggregate the local main system behavior models trained by one or more other spectrum management devices as various federated learning clients participating in the federated learning to obtain the final global model as the main system behavior model. In this example, the spectrum management device, i.e., the electronic device 100, can itself participate in the federated learning as a federated learning server to obtain the main system behavior model. For example, the electronic device 100 can replace NWDAF_0 shown in FIG. 2(a) as a federated learning server, and one or more other spectrum management devices can replace NWDAF_1 to NWDAF_4 shown in FIG. 2(a) as federated learning clients, which will not be repeated here.
[0057] As an example, the main system behavior model obtaining unit 103 may be configured to send a federated learning request to other spectrum management devices in the federated prediction area associated with it, and receive a federated learning response from other spectrum management devices that agree to participate in the federated learning. As an example, the federated learning response includes the number of geographical units occupied by user devices in the secondary system associated with the federated learning client.
[0058] As an example, the main system behavior model obtaining unit 103 may be configured to: in each round of federated learning training, aggregate the local main system behavior models trained by each federated learning client based on the weight to obtain a global model, and send the global model to each federated learning client, so that each federated learning client performs the next round of training based on the global model; and until a predetermined condition is met, stop the federated learning, and use the obtained final global model as the main system behavior model. As an example, the predetermined condition includes that the final global model reaches a predetermined accuracy and / or has been trained for a predetermined number of rounds.
[0059] As an example, the spectrum management device, i.e., the electronic device 100, can be used as a federated learning client, and the main system behavior model obtaining unit 103 can be configured to perform federated learning together with one or more other spectrum management devices that are federated learning clients, and aggregating the local main system behavior models trained by each federated learning client to obtain a global model as the main system behavior model. For example, the spectrum management device, i.e., the electronic device 100, serves as a federated learning client in distributed federated learning.
[0060] As an example, the spectrum management apparatus, ie, the electronic device 100, is implemented by a base station in the secondary system or an edge server (edge computing node) associated with the secondary system.
[0061] Fig. 2(b) shows another example of the system structure according to the embodiment of the present disclosure. It should be noted that the system structure according to the embodiment of the present disclosure is not limited to Fig. 2(a) and Fig. 2(b), but there are various other variations.
[0062] Figure 2(b) shows a distributed federated learning solution based on edge computing, in which the local training and model aggregation of federated learning are undertaken by base stations or edge servers.
[0063] The system in Figure 2(b) consists of an access layer and a data layer. The bottom layer is the access layer, which represents the access relationship between the user equipment and the base station. The upper layer in the system structure is an abstract data layer, which is composed of, for example, the edge server of the base station. In Figure 2(b), there is one PU and multiple SUs. The base station in the PU is represented by BS. The base station set of multiple SUs is N = {BS1, BS2, BS3, BS4, BS5, BS6, BS7, BS8}. For example and not limitation, the base station set is divided into two federated prediction zones FPZ (Federated Prediction Zone): Federated Prediction Zone 1, namely FPZ_1, and Federated Prediction Zone 2, namely FPZ_2, where FPZ_1 = {BS1, BS2, BS3, BS4}, and FPZ_2 = {BS5, BS6, BS7, BS8}.
[0064] As shown in Figure 2(b), for example, 8 edge computing nodes Edge_1 to Edge_8 are included, which are associated with BS1 to BS8 respectively. For example, FPZ_1 includes Edge_1, Edge_2, Edge_3, and Edge_4, each edge computing node trains a local model using data obtained from an associated base station as a federated learning client, and one edge computing node (e.g., Edge_2) among Edge_1, Edge_2, Edge_3, and Edge_4 acts as a federated learning server to globally aggregate the local model to obtain a final global model (e.g., global model a). For example, FPZ_2 includes Edge_5, Edge_6, Edge_7, and Edge_8. Each edge computing node acts as a federated learning client to train a local model using data obtained from an associated base station. One edge computing node (e.g., Edge_6) among Edge_5, Edge_6, Edge_7, and Edge_8 acts as a federated learning server to globally aggregate the local model to obtain the final global model (e.g., global model b).
[0065] The electronic device 100 may be Edge_2, that is, the electronic device 100 obtains the global model a for FPZ_1 as the main system behavior model for FPZ_1; or the electronic device 100 may be Edge_6, that is, the electronic device 100 obtains the global model b for FPZ_2 as the main system behavior model for FPZ_2. For the convenience of description, the edge server is also referred to as a spectrum management device, however, the electronic device 100 such as Edge_2 or Edge_6 may not have functions such as spectrum allocation and / or interference analysis.
[0066] As an example, when the spectrum management device, i.e., the electronic device 100, serves as a federated learning client in distributed federated learning, the main system behavior model obtaining unit 103 may be configured to send the number of geographical units occupied by user devices in the subsystem associated with the spectrum management device to other federated learning clients. For example, Edge_2 shown in FIG2(b) sends the number of geographical units occupied by user devices in the subsystem associated with Edge_2 to Edge_1, Edge_3, and Edge_4.
[0067] As an example, the main system behavior model obtaining unit 103 may be configured to, in each round of training of federated learning, aggregate the local main system behavior model trained by the spectrum management device and the local main system behavior model received from other federated learning clients through broadcasting based on weights to obtain a global model (for example, Edge_2 shown in FIG2(b) aggregates the local main system behavior model of Edge_2 and the local main system behavior models received from Edge_1, Edge_3, and Edge_4 through broadcasting based on weights to obtain a global model), and use the global model as the new local main system behavior model for the next round of training; and until a predetermined condition is met, stop federated learning, and use the obtained final global model as the main system behavior model. As an example, the predetermined condition includes that the final global model reaches a predetermined accuracy and / or a predetermined number of rounds of training have been performed.
[0068] As an example, a plurality of federated prediction areas are constructed based on the geographical locations of base stations in the secondary system, and the size of the federated prediction area is embodied by the number of geographical cells having a predetermined geographical size.
[0069] Figure 3 FIG. 2 is an example showing how to divide a network according to an embodiment of the present disclosure.
[0070] like Figure 3 As shown, the network is divided into geographical units (also called minimum units or cells). For example but not limitation, the geographical unit can be a square cell of 100m×100m. The cells are numbered, such as Figure 3As shown, the network is divided into the minimum unit set U = {0, 1, 2, ..., 63}.
[0071] For example, Figure 3 As shown in the example, the PU is located in the cell numbered 22, and the federal prediction zone FPZ1 is a 3×3 cell, namely, it includes cells numbered 0, 1, 2, 8, 9, 10, 16, 17, and 18.
[0072] As an example, the number and location of the plurality of federated prediction regions are determined by a k-means algorithm and a density-based spatial clustering algorithm DBSCAN for noisy applications. The above combination of the k-means algorithm and DBSCAN may be referred to as a K-DBSCAN algorithm.
[0073] As an example, the k-means algorithm is used to cluster base stations in a secondary system, delete clusters that do not contain any base stations without changing the position of the cluster center, and set the position of a specific base station in the cluster as a new cluster center when the distance between the center of the cluster to which it belongs exceeds a first predetermined distance threshold. The DBSCAN algorithm uses the position of the cluster center determined by the k-means algorithm as input to obtain a new cluster center, merges the cluster centers among the new cluster centers whose distances are less than a second predetermined distance threshold, and repeats the k-means algorithm and the DBSCAN algorithm until the number of cluster centers no longer changes, wherein the cluster center serves as the center of the federal prediction area and the number of cluster centers corresponds to the number of multiple federal prediction areas.
[0074] Figure 4 is an example of a pseudo code illustrating a K-DBSCAN algorithm according to an embodiment of the present disclosure.
[0075] As an example, the K-DBSCAN algorithm first sets the initial clustering K value of the network. The initial K value should be set to the square of an integer to facilitate the determination of the location of the K cluster centers. Since this value is not the number of the final federal prediction areas, a larger value can be set at the beginning. The initial K cluster centers are evenly distributed in the network. First, the k-means algorithm is run to perform the initial clustering of the base stations. When the location of the cluster center does not change, the cluster that does not contain the base station is deleted. Secondly, when the distance between a base station in the cluster and its cluster center exceeds the predetermined value D max After that, the base station location is set as the new cluster center. At this point, the cluster center location is used as input to run the DBSCAN algorithm to obtain a new cluster, and finally the new cluster centers are traversed to merge the cluster centers with similar distances as the new cluster center. The distance is, for example, the Euclidean distance.
[0076] Figure 4The pseudo code of is described in conjunction with NWDAF. More specifically, in Figure 4 In pseudo code:
[0077] Input: Coordinates (x, y) of N base stations in the secondary system: N = {(x1, y1), ..., (x i ,y i ),...,(x N ,y N )};
[0078] Output: cluster centers K and their associated base stations;
[0079] 1: Initialization: Set the maximum radius D of the cluster max . Set the initial value of K and initialize K cluster centers to be evenly distributed in the network.
[0080] K={(X1,Y1),...,(X i , Y i ),...,(X K , Y K )};
[0081] 2: Repeat steps 3-11 below
[0082] 3: Run the k-means algorithm with the input as the preset K cluster centers K and the location of the base station N
[0083] 4: Delete empty cluster centers in K
[0084] 5: For all (x in N n ,y n )conduct
[0085] 6: if(x n ,y n ) and its cluster center has a Euclidean distance greater than or equal to D max ,So
[0086] 7: Add new cluster centers to N (x n ,y n );
[0087] 8: End if
[0088] 9: End for
[0089] 10: Run the DBSCAN algorithm with the input as the cluster center K
[0090] 11: Merge cluster centers in K if their distance is close to 0
[0091] 12: Until the number of clusters remains unchanged
[0092] Traditional clustering algorithms such as k-means are limited by the setting of the initial K value. When the K value is determined, the number of clusters is also a constant value, so the number of federated prediction areas cannot be adjusted adaptively. In addition, the k-means clustering algorithm can only process spherical clusters, that is, a solid cluster (this is because the algorithm itself is limited in calculating the average distance). But there are often various shapes in reality, such as rings and irregular shapes. In these cases, clustering algorithms based on high sample density are applicable. The DBSCAN density-based clustering algorithm does not have a cluster center point, so it will cause irregular cluster shapes. After clustering, the distance between base stations located at the edge of the cluster may be far, and when the base station density is high, it cannot output effective clusters or the number of clusters is too low. The K-DBSCAN algorithm retains the regular shape characteristics of the k-means algorithm after clustering, and avoids the defect of being limited by the initial K value setting, and can adaptively adjust the number of clusters. Its purpose is that the federated prediction area formed after clustering is composed of base stations around the cluster center, so that the main system received signal strength information provided by the base station can be fully fed into the learning model. Therefore, the K-DBSCAN algorithm can optimize the number of clusters and thus reduce the number of federated prediction regions.
[0093] Figure 5(a) and 5(b) It is an example showing the effect of using the K-DBSCAN algorithm to determine the federal prediction area according to an embodiment of the present disclosure. As shown in Figure 5(a), the network size is set to 3km×3km, the number of base stations is 300, and the dots in Figure 5(a) represent the distribution of base stations. Figure 5(b) shows the effect after clustering using the K-DBSCAN algorithm, where the asterisks represent the locations of the centers of the 82 federal prediction areas (marked as Centers in the figure), and the dots represent the distribution of base stations (marked as Base Stations in the figure). It can be seen from Figure 5(b) that the two-layer clustering algorithm (K-DBSCAN algorithm) can minimize the number of federal prediction areas, so that local data can be fully utilized in federated learning.
[0094] The way of obtaining the federal prediction area in the embodiments of the present disclosure is different from the user clustering in the existing federated learning. The user clustering in the existing federated learning is to cluster and group users with similar data, select representative user data from the group for training, and then realize the sampling processing of the data. The key point is that the clustering of users in the existing federated learning technology does not affect the structure of the learning model in the federated learning, but only screens the training data. In the embodiments of the present disclosure, for example, clustering is performed according to the geographical location distribution of the base stations, and the size and location of the clustered federal prediction area will determine the input and output targets of the model in the federated learning. The federal prediction area decouples the main system signal prediction problem in a large-scale network into prediction problems of multiple sub-areas, which reduces the communication overhead of centralized prediction data collection on the one hand, and reduces the computational complexity of model training on the other hand.
[0095] Figure 6 FIG. 1 is an example of a process for constructing a federated prediction region according to an embodiment of the present disclosure. Figure 6 In the example, the system structure of Figure 2(a) is used to construct the federated prediction region. The federated learning in Figure 2(a) is centralized federated learning.
[0096] In A1, if Figure 3 The network is divided into geographical units as shown. In A2, the electronic device 100 collects the geographical location information of the base station, and determines the number and location of the federal prediction areas based on the location information uploaded by the base station, for example, by running the K-DBSCAN dual clustering algorithm (it should be added here that when the electronic device 100 directly participates in federated learning as a federated learning server, the electronic device 100 also collects the geographical location information of the base station, and determines the number and location of the federal prediction areas based on the location information uploaded by the base station, for example, by running the K-DBSCAN dual clustering algorithm). In A3, after determining the division of the federal prediction area, NWDAF needs to register the federal learning information in the network. As shown in the example of Figure 2(a), the network is divided into two federal prediction areas, FPZ_1 and FPZ_2. When the base station associated with NWDAF is divided into the federal prediction area, NWDAF will register with NRF as a federal learning client for the federal prediction area. The NWDAF that undertakes the task of aggregating the federated learning model is registered as a federated learning server. As shown in Figure 2(a), the federated prediction zone 1, i.e., FPZ_1, contains {NWDAF_1, NWDAF_2}, the federated prediction zone 2, i.e., FPZ_2, contains {NWDAF_3, NWDAF_4}, and NWDAF_0 is registered as FL-Server. In A4, after the federated learning information is registered, the NWDAF corresponding to the federated prediction zone constructs the federated prediction zone members (forms a federated learning group).
[0097] Figure 7It is a schematic diagram showing the information interaction for constructing a federated learning group according to an embodiment of the present disclosure. It is described in conjunction with Figure 2(a). First, the NRF notifies the NWDAF (for example, NWDAF_0 in Figure 2(a)) that serves as the federated learning aggregation point to become the FL-Server of the federated prediction area through FL_Zone Notify. This information will also include the identity information of other FL-Clients (for example, NWDAF_1 to NWDAF_4 in Figure 2(a)) in the current federated prediction area. The FL-Server requests the FL-Client to participate in the federated prediction area through NWDAF_FL_Zone Request. If the FL-Client confirms to participate in the federated learning, the FL-Client sends a NWDAF_FL_Zone Response to the FL-Server for confirmation. The NWDAF_FL_ZoneResponse information can also include the number of geographical units occupied by user devices in the subsystem corresponding to each FL-Client. The purpose is to provide guidance for the subsequent federated learning model aggregation, for example, to calculate the aggregation weight of the local model based on this number.
[0098] Taking the system structure of Figure 2(b) as an example, the process of constructing a federal prediction area is as follows. B1. Divide the geographical units. B2. For example, run the K-DBSCAN algorithm to cluster the base stations and divide the federal prediction areas. In the distributed federated learning network shown in Figure 2(b), the federal prediction areas can be divided by the network management OAM (Orchestration and management) or another centralized spectrum management device. B3. After the federal prediction areas are divided, the base stations or edge servers in the same federal prediction area form a federated learning group. As shown in Figure 2(b), federal prediction area 1, namely FPZ_1, contains: {Edge_1, Edge_2, Edge_3, Edge_4}, and federal prediction area 2, namely FPZ_2, contains: {Edge_5, Edge_6, Edge_7, Edge_8}.
[0099] In order to optimize the computing energy consumption and communication energy consumption in the federated learning model, according to the requirements of model accuracy in different specific scenarios, the overall energy consumption of the model is reduced as much as possible while meeting the accuracy. Among them, the computing energy consumption comes from local model training, and the communication energy consumption comes from model parameter aggregation. Since the two parts of energy consumption are relatively independent in the entire model training environment, they can be optimized separately, thereby reducing the difficulty of solving the optimization problem. Among them, in the optimization of computing energy consumption of federated learning, the energy efficiency (accuracy / energy consumption) of federated learning can be optimized when convergence is achieved by determining a better range for each federated prediction area (for example, a rectangle with the minimum grid unit as the unit); in the optimization of communication energy consumption of distributed federated learning, the location of the federated aggregation point for local model aggregation (also called model aggregation) can be selected according to the real-time channel state (that is, a federated learning client (hereinafter, also referred to as an access point (AP)) is selected for local model aggregation) to minimize the overall communication energy consumption, thereby improving the overall energy efficiency of federated learning.
[0100] An example of determining a preferred range for each federal prediction area is described below.
[0101] As an example, lower bound values of the sizes of the plurality of federated prediction regions are determined based on the distance between the center of the federated prediction region and the secondary system, and each federated prediction region is updated based on the accuracy and / or energy consumption when the primary system behavior model converges.
[0102] Accuracy and / or energy consumption may be collectively referred to as energy efficiency.
[0103] The method steps for determining the federal forecast area are as follows.
[0104] 1.1 Use geometric methods to determine the lower bound of the size of each federated prediction region.
[0105] 1.1.1 Obtain the center of each federated prediction area and the locations of the base stations it contains based on the K-DBSCAN algorithm or the k-means clustering algorithm.
[0106] 1.1.2 Based on the horizontal and vertical coordinates, calculate the maximum distance from the center point among all base stations, recorded as Δx and Δy respectively.
[0107] 1.1.3 Take the maximum value of Δx and Δy, round it up and multiply it by 2 as the minimum radius of the federal prediction area.
[0108] 1.2 For each federated prediction area, based on its lower bound value, the idea of finding the extreme value of the function is used by the three-point method in calculus to train three federated learning models (local models) in parallel, evaluate the energy efficiency of the three models after convergence, and then use the "three-point method" to update the parameters and determine the learning models of three different dimensions. The main steps of the three-point method are as follows:
[0109] 1.2.1 Based on the minimum radius r0, take r1 = r0 + s and r2 = r0 + 2*s, and use r0, r1, and r2 as the radii corresponding to the federated learning models of three input dimensions respectively (where s is the step factor).
[0110] 1.2.2 Train the three models respectively under the same environment, and calculate the energy efficiency when the three models converge, corresponding to E0, E1, and E2 respectively.
[0111] 1.2.3 Update r0, r1, and r2 according to different situations, and the updates are as follows:
[0112] (1) E0 < E1 < E2, r0 = r1, r1 = r2, r2 = r2 + s
[0113] (2) E0 > E1 > E2, and r0 - s > the minimum radius, r0 = r0 - s, r1 = r0, r2 = r1
[0114] (3) E0 > E1 > E2, and r0 - s < the minimum radius, and s ≤ 1, the optimal radius = r0,
[0115] (4) E0 < E2 < E1, and s ≥ 2, s = s / 2, r0 = r1, r1 = r0 + s
[0116] (5) E2 < E1 < E1, and s ≥ 2, s = s / 2, r0 = r1, r1 = r2 - s, r2 = r1
[0117] (6) For the relationships of E0, E1, and E2 in (4) and (5), if s < 2, then output the radius corresponding to the maximum energy efficiency.
[0118] 1.2.4 Repeat 1.2.2 - 1.2.3 several times until the predetermined conditions are met.
[0119] 1.3 Output the optimal radius of each federated prediction area.
[0120] It should be noted that the optimal area may not be a circular area either.
[0121] Examples of selecting federated aggregation points (which can also be called model aggregation points or aggregation points) for local model aggregation in distributed federated learning are described below.
[0122] As an example, in the case where the spectrum management device, i.e., the electronic device 100, acts as a federated learning client in distributed federated learning, the spectrum management device, i.e., the electronic device 100, is selected for aggregation under the following predetermined conditions: based on the channel information between the spectrum management device and one or more other spectrum management devices, the sum of the transmission energy consumption of the transmission path between the spectrum management device and one or more other spectrum management devices is calculated when the spectrum management device acts as an aggregation device for aggregating local main user behavior models within a federated prediction area associated with it, wherein the sum of the transmission energy consumption when the spectrum management device acts as an aggregation device is less than the sum of the transmission energy consumption when any other spectrum management device of one or more other spectrum management devices acts as an aggregation device.
[0123] As an example, the unit 103 for obtaining the primary system behavior model may be configured to exchange channel information with one or more other spectrum management devices through broadcasting.
[0124] As an example, the transmission energy consumption of the transmission path is calculated based on the transmission rate and transmission power of the transmission path.
[0125] Here, the spectrum management device and one or more other spectrum management devices are collectively referred to as APs for explanation. For all APs in each federal prediction area (cluster), the channel state information is shared. Each AP is selected in turn, and the AP calculates the communication energy consumption of other APs when the model is aggregated at the AP, and the AP records the result. After all APs have calculated the aggregated energy consumption when they are used as aggregation points (aggregation devices, aggregation centers), all APs share energy consumption information with each other, and then select the AP with the minimum energy consumption as the model aggregation point. In other words, when determining the aggregation center, APs exchange channel information with each other by broadcasting, calculate the overall communication energy consumption of different APs as aggregation centers based on the channel information, and select the AP corresponding to the minimum communication energy consumption as the federated learning aggregation center.
[0126] Figure 8 An example of selecting a federation aggregation point for performing local model aggregation according to an embodiment of the present disclosure is shown.
[0127] like Figure 8 As shown in the figure, if AP1 is selected as the aggregation point, there are three wireless transmission paths between AP1 and other APs, namely AP2-AP4. The transmission rates corresponding to the three wireless transmission paths can be calculated based on the free space loss model and the Rayleigh fading model, and the transmission energy consumption can be calculated based on the transmission power. The sum of the transmission energy consumption corresponding to the three wireless transmission paths is less than the sum of the corresponding transmission energy consumption when any AP among AP2-AP4 is used as an aggregation device.
[0128] Consider a time slot system, that is, at the beginning of each time slot, the terminal (user equipment) of the secondary system reports the received PU signal power to its base station. The goal of using the primary system behavior model is to predict the future signal reception strength based on the historical PU received signal strength.
[0129] As an example, the main system behavior model is a multi-input and multi-output long short-term memory neural network model (LSTM model for short).
[0130] The use of single-input single-output LSTM will increase the number of models in the system, which will lead to an increase in communication overhead. In the federated aggregation process of the multi-input multi-output LSTM used in this application, the federated learning client (aggregation client) only needs to transmit a single local model, and does not need to establish communication connections with the device or function used for model aggregation multiple times, thereby reducing communication overhead.
[0131] As an example, the dimensions of the input and output of the MIMO LSTM neural network model are determined by the size of the federated prediction region.
[0132] As an example, for each federal prediction area, the input of the multi-input multi-output long-term and short-term neural network model is the main system signal reception power information of all geographical units included in the federal prediction area in a predetermined number of historical time slots, and the output is the predicted value of the main system signal reception power of each geographical unit in the next time slot.
[0133] Although the main system signal receiving power is different between geographical units at different locations, the signal change trend of the main system is consistent. Therefore, using the main system signal receiving power information of multiple geographical units as the input of the LSTM model at the same time can better learn the power change of the main system.
[0134] As an example, the main system behavior model is obtained by aggregating local main system behavior models obtained from federated learning clients participating in the federated learning.
[0135] For example, the federated learning client obtains the local main system behavior model by training the local data. The main system behavior model can be obtained by aggregating the local main system behavior model.
[0136] As an example, the local primary system behavior model corresponding to the federated learning client may be aggregated by calculating a weight based on the number of geographical units occupied by user devices in the secondary system associated with the federated learning client.
[0137] Assume that the federated prediction area contains N FL-Clients, and the number set of SU-related SUs of N FL-Clients in the federated prediction area is S = {S1, S2, ..., S n, ..., S N}, where S n represents the number of geographical units occupied by the user devices in the secondary system associated with the nth FL-Client. Assume that the local model parameter set of FL-Client is in, Represents the local model parameters of the nth FL-Client. Assume that the number of geographical units in the federated prediction area is M.
[0138] The weight value set of the model can be calculated by formula (1) A = {α1, α2, ..., α n , ..., α N}:
[0139]
[0140] Among them, S n / M is to normalize the number of geographical units occupied by the user equipment in the FL-Client and its associated secondary system. After normalization, the exponential function can protect the local main system behavior model with sparse SU distribution and avoid its weight value being too low in aggregation. After calculating the weight value, the local model can be aggregated based on the weight value. For example, the local model parameters are multiplied by the weight value and then summed to obtain the parameters of the global model.
[0141] As an example, when a federated learning client transmits a local model to a device for model aggregation (e.g., a federated learning server, or a federated learning client for model aggregation in distributed federated learning), the number of geographical units occupied by user devices in the associated subsystem is transmitted, instead of the specific location of the user devices in the associated subsystem. Location information is private information, and disclosure of the location may lead to the risk of attack or eavesdropping. In an embodiment according to the present disclosure, providing the specific location of the user device in the SU is avoided, thereby protecting the privacy of the user. That is, during the model aggregation process, the FL Client will not disclose the location information of the user device in the current SU, but only transmits the number of geographical units occupied by the user devices in the associated subsystem, thereby ensuring the security of the location privacy information of the SU.
[0142] As an example, the local main system behavior model corresponding to the federated learning client can be aggregated by calculating the weight based on the distribution status of the user devices in the secondary system associated with the federated learning client. The aggregation method based on the weight according to the embodiment of the present disclosure can effectively filter the local model, thereby improving the accuracy of the federated learning model.
[0143] For example, the aggregation weight of the FL client can be calculated based on the distribution of user devices in the secondary system. When the geographical locations of user devices are relatively concentrated (since users are concentrated in a few locations and there are no users in other locations, only the power information of the primary users in a few locations can be perceived), only accurate predictions of the power of the primary users in the local area can be provided, and the aggregation weight of the FL client is relatively low; when the distribution of user devices is relatively uniform (since users in different locations are evenly present, the power information of the primary users in more locations can be perceived, and the signal strength of the primary users in more locations can be obtained, so there is better quality training data when training the local model), the FL client has a higher aggregation weight.
[0144] In other words, the federated learning scheme uses a weighted model aggregation, where the weight is calculated by the distribution of users served by the base stations in the secondary system. When the users served by the base station are evenly distributed in the federated prediction area, the local model provided by the base station has a higher weight. The principle is that base stations with evenly distributed users can obtain the signal strength of primary users in more locations, so the base station has better training data when training the local model. When performing global model aggregation, the network parameters of the local model will form the network parameters of the global model in a weighted form.
[0145] As an example, the local main system behavior models may be aggregated by calculating weights based on the model accuracy of the local main system behavior models corresponding to the federated learning clients.
[0146] For all federated learning clients (APs) in a federated prediction area, the user data owned by each AP may be different, so different weights need to be assigned to the models trained by different APs when performing federated aggregation. The weight distribution scheme can adopt a distribution method based on model accuracy, using direct or indirect accuracy expressions to determine the weight of each model.
[0147] As an example, the model accuracy of the local main system behavior model increases as the amount of user data involved in the local main system behavior model increases, and decreases as the loss of the local main system behavior model increases. Based on the aggregation method of weight allocation based on model accuracy according to the embodiment of the present disclosure, the model training speed is improved and the energy consumption overhead of the model reaching convergence is reduced.
[0148] For example, the definition of the accuracy of the deep learning model as the main system behavior model can be: the ratio of the number of correct data to the number of all data, but this definition is not applicable in actual use, so it is necessary to design an expression that indirectly represents the accuracy of the model: for example, a function of the amount of user data and model loss can be designed, and its value increases with the amount of user data and decreases with the model loss. Compared with the general average weight distribution method, this aggregation weight determination method has a higher weight for the local model with higher model accuracy during aggregation. After aggregation, the federated learning model can converge faster, reduce the rounds of aggregation communication at the communication level, and thus reduce communication energy consumption.
[0149] For example, the weight calculation design of each local model is as follows: the logarithmic value of the amount of user data is used as the numerator, and the model loss is used as the denominator. Using the logarithmic function of the amount of user data as the numerator not only ensures its monotonically increasing property, but also reasonably reflects the trend of the impact of the amount of user data on the model accuracy; while using the model loss as the denominator, while ensuring its monotonically decreasing property, fully reflects the important impact of the local model fitting effect on the overall effect of the federated learning model.
[0150] As an example, when determining the aggregation weight, the following formula (2) may be used:
[0151]
[0152] Where, Indicator(n)=log2(1+d n ) / η n
[0153] In the above two equations, ω n represents the weight of the local model of the nth AP in the federated prediction area, |C i | represents the number of APs in the federated prediction area, d n represents the amount of user data owned by the nth AP, η n represents the loss of the local model of the nth AP.
[0154] For the same model, the larger the federation aggregation period (model aggregation period), the faster the convergence speed, but the loss will be higher when reaching convergence; the smaller the federation aggregation period, the slower the convergence speed, but the loss will be lower when reaching convergence. Based on this, in the overall energy consumption optimization of federated learning, a function about the federation aggregation period can be designed, and this function is used to calculate and determine the period of each round of aggregation. As an example, the energy consumption of communication calculation under different aggregation periods can be calculated to determine the aggregation period that minimizes energy consumption.
[0155] In an embodiment according to the present disclosure, a function about the federation aggregation period is designed, which decreases monotonically over time. By continuously adjusting the aggregation period based on the function about the federation aggregation period, a compromise is made between the accuracy and convergence speed of the federated learning model, and the communication rounds of parameter aggregation are reduced while ensuring a certain model accuracy, thereby reducing the communication energy consumption of model training.
[0156] As an example, as the loss of the primary system behavior model decreases monotonically, the period of the aggregation also decreases monotonically.
[0157] As an example, if the loss of the local main system behavior model is not monotonically decreasing, the period of aggregation is reduced by a predetermined ratio.
[0158] When determining the aggregation period, each federated prediction region records the aggregation period of the previous time slot and the loss at the current moment. Considering the possible fluctuations in loss during model training, when the loss shows a monotonically decreasing trend, the model aggregation period is updated using the designed monotonically decreasing function; in other cases, the model aggregation period is reduced by a fixed ratio.
[0159]
[0160]
[0161]
[0162] Among them, F(x) in formula (3) represents the average loss of the federated model, f(x; s i ) is the loss of each local model in federated learning, x is the local model parameter, and s is the training dataset.
[0163] Formula (4) represents the update iteration of model parameters, where the upper branch in formula (4) represents the update of the federated aggregation mode, k mod τ = 0 means that the federated aggregation time node has been reached, so all local models need to be aggregated and the aggregation results are sent to each local model to complete the parameter update, where τ represents the federated aggregation period; the lower branch in formula (4) represents the independent update of the local model, k mod τ ≠ 0 means that the federated aggregation time node has not been reached, and each local model only updates its parameters based on its own training results, where g() represents the stochastic gradient, η k is the preset parameter.
[0164] Formula (5) represents the update iteration of the federation aggregation cycle, where T0 is the preset time parameter, τ0 is the initial federation aggregation cycle, and l represents a positive integer greater than 1. The two parts in formula (5) represent the update rules of the federation aggregation cycle under different conditions: the upper branch represents the general situation, which assumes that the federation learning model will be trained in an overall positive trend, that is, the model loss will gradually decrease, and the federation aggregation cycle will also continue to decrease; the lower branch represents the special situation, which takes into account the fluctuations that may occur during the training of the federation learning model, that is, a certain update will make the federation learning model go in a bad direction, but in order to ensure the final effect of the model, it is still necessary to reduce the federation aggregation cycle. It should be noted that for special cases, 1 / 2 is only a reference value. In theory, any value set between (0,1) is acceptable.
[0165] The LSTM model network parameters are set so that the input of the model is T×M dimensional, where T is the above-mentioned predetermined number of historical time slots, and M is the number of all geographic units included in the corresponding federal prediction area.
[0166] When the function of model training and aggregation of federated learning is in the core network (e.g., NWDAF) (as shown in Figure 2(a)), since the core network has sufficient computing power, there is no need to consider the computational overhead of federated learning when training the federated learning model. Therefore, no upper limit is set for the input T (predetermined number of historical time slots) of the LSTM model.
[0167] In contrast, when the model training and aggregation functions of federated learning are implemented by the edge computing node (edge server) of the base station (for example, as shown in Figure 2(b)), the edge server generally has limited computing power and must also process computing tasks in the access network. Therefore, its computing overhead needs to be considered in federated learning.
[0168] When the computational overhead is considered, the prediction of the primary user received power under the system structure shown in FIG2( b ) mainly includes the following steps T1 to T5 .
[0169] First, we describe step T1. The input and output dimensions of the LSTM model are determined according to the size of the federated prediction region, such as Figure 3 In the example shown, the federated prediction area 1 consists of a 3×3 minimum cell area. At this time, the input of the LSTM model needs to contain the PU signal receiving power information of 9 minimum cells. To confirm the input and output dimensions of the model, the computing power of the FL-Client needs to be considered. Let the average computing power of the FL-Client be C, in cycles / second, that is, the number of CPU cycles that can be executed per second. Let the user's local training time in each time slot be τ, and the number of iterations of local training be γ.
[0170] Fig. 9 is an example showing the structure of an LSTM network model according to an embodiment of the present disclosure. Fig. 9 In the above equation, T represents the number of input historical time slots and S represents the size of the federated prediction region ( Figure 3 In the example shown, S = 9), none is an intermediate parameter during model operation, hidden_size1 and hidden_size2 are the number of hidden nodes in the two-layer LSTM, and pred_time is the number of time slots predicted in the future.
[0171] The computational complexity of the LSTM layer is analyzed. LSTM mainly consists of input gate, forget gate, output gate and cell state update. Taking time t as an example, the input gate inputt is calculated as input t =σ(W ii x t +b ii +W hi h t-1 +b hi ). Where σ is the Sigmoid activation function, W ii is the input gate network weight value, and its dimension is (H, T×S), where H is the number of hidden nodes. t is the input, with a dimension of T×S. ii is the bias value, and its dimension is H. hi To retain the network weight value of the hidden state input at the previous moment, the dimension is (H, T×S). t-1 is the hidden state at the previous moment, with a dimension of H. b hi is the bias value, and its dimension is H. The matrix multiplication and addition operations inside the activation function can be calculated to have a complexity of O(H×T×S+2H+H 2 ), when using the Relu activation function, the input gate calculation complexity is O(4(H×T×S+2H+H 2 )). Similarly, calculate the forget gate forget t =σ(W if x t +b if +W hf h t-1 +b hf ) and the output gate out t =σ(W io x t +b io +W ho h t-1 +b ho ), cell state update g t =tanh(W ig x t +b ig +W hgh t-1 +b hg ), the final LSTM output hidden state h t =out t ⊙c t-1 , cell state c t =forget t ⊙c t-1 +input t ⊙g t The computational complexity of a single-layer LSTM is O(4(H×T×S+3H+H 2 )). Similarly, the second LSTM layer, LSTM2, uses the input of the first LSTM layer, LSTM1, as the output for operation. The computational complexity of the two-layer LSTM is O(4(H1×T×S+3H1+H1 2 )+4(H2×H1+3H2+H2 2 )), where H1 and H2 are hidden_size1 and hidden_size2, i.e. the number of nodes in the hidden layer. Finally, a fully connected layer (Dense) is added. Therefore, the complexity of the network's forward propagation is O f =O(4(H1×T×S+3H1+H1 2 )+4(H2×H1+3H2+H2 2 )+H2(T pred ×S+1)). The network has the same complexity when performing back propagation as forward propagation, so the final complexity is O train =2O f According to the calculated complexity, the input history time slot length of the current FL-Client that meets the time limit requirement can be obtained, that is, Where α is the computation density, that is, the number of CPU cycles required to perform one operation. When the maximum delay is used as the training time, the input historical time slot length can be inferred
[0172] Next, step T2 is described. A LSTM deep neural network based on Tensorflow is constructed based on the parameters calculated in the previous step T1. For example, a deep neural network constructed with T=10 can be used, that is, 10 historical time slots are used to predict the PU received power of the next time slot, combined with Figure 3In the example shown, S=9, the input dimension of the LSTM deep neural network is 10×9=90, and the output dimension is 9. The input and output dimensions in this embodiment are just an example of a case. When the federated prediction area changes, the input and output dimensions should also be adjusted accordingly. For example, when the federated prediction area contains a minimum number of cells of M and the prediction time slot length is pred_time, the input dimension is T×M and the output dimension is M×pred_time.
[0173] Assume that the network structure and parameters of LSTM are as follows:
[0174] The first layer is a standard LSTM layer implemented in Keras (an open source artificial neural network library), with an input data dimension of (9, 10), 128 hidden nodes, and a sigmoid activation function;
[0175] The second layer is a standard LSTM layer implemented by Keras, with 64 hidden nodes and the activation function is the Sigmoid function;
[0176] The third layer is the fully connected layer.
[0177] Next, we describe step T3. After the LSTM deep neural network is constructed, FL-Client trains the model based on local data. For example, the loss function of model training is defined as the mean square error between the PU predicted received power and the actual received power. When the local training converges or reaches the number of iterations, FL-Client broadcasts the model parameters in the federated prediction area. For example, as shown in Figure 2(b), Edge_1 in FPZ_1 will pass the trained local model parameters to {Edge_2, Edge_3, Edge_4}, and similarly, other edge computing nodes will also pass the model parameters to Edge_1.
[0178] Next, step T4 is described. After receiving the local model parameters of all the federated prediction areas, the FL-Client performs weighted aggregation on the model parameters. The weights may be determined based on at least one of formula (1), formula (2), or the distribution state of user devices in the subsystem associated with the federated learning client.
[0179] Finally, step T5 is described. FL-Client uses the aggregated global model as the local model to perform a new round of training based on local data, and repeats steps T3-T5 until a predetermined condition is met, where the predetermined condition includes that the global model reaches a predetermined accuracy or has been trained for a predetermined number of rounds.
[0180] The prediction of the primary user received power under the system structure shown in FIG. 2( a ) mainly includes the following steps S1 to S5 .
[0181] First, step S1 is described. The LSTM model network parameters are set. The input of the model is T×M dimensional, where T is the number of historical time slots. Since the core network server has sufficient computing power, there is no upper limit on the T value. M is the minimum number of cells corresponding to the federated prediction area.
[0182] Then, step S2 is described. Similar to step T2 described above in conjunction with FIG. 2( b ), an LSTM deep neural network is constructed.
[0183] Let's describe step S3. The NWDAF acting as the FL-Client will collect data uploaded by the base stations under its jurisdiction. As shown in Figure 2(a), there are four base stations in the federal prediction area 1, FPZ_1, and the data sets collected by the base stations are {Dataset 1, Dataset 2, Dataset 3, Dataset 4}. For example, the base stations under the jurisdiction of NWDAF_1 include {BS1, BS2}, so the collected data is {Dataset 1, Dataset 2}. Since base stations may have overlapping coverage areas, the FL-Client needs to preprocess the collected data sets. The specific preprocessing method is that when there are multiple received powers of the main system received signals corresponding to the same minimum cell in the data set, the received powers of the multiple main system signals are averaged.
[0184] Next, step S4 is described. After the FL-Client training model converges or reaches the aggregation period, the FL-Client uploads the model parameters to the NWDAF that serves as the FL-Server. As shown in Figure 2(a), NWDAF_0 collects the local model parameters uploaded by NWDAF_1 and NWDAF_2 in FPZ_1, and collects the local model parameters uploaded by NWDAF_3 and NWDAF_4 in FPZ_2. FL-Server calculates the model aggregation weight value for FPZ_1 and the model aggregation weight value for FPZ_2. The weight value calculation method is the same as step T4 described in conjunction with Figure 2(b).
[0185] Finally, step S5 is described. FL-Server sends the aggregated global model to FL-Client, and FL-Client uses the global model as a new local model for a new round of training, repeating steps S3-S5 until a predetermined condition is met, where the predetermined condition includes that the global model reaches a predetermined accuracy or has been trained for a predetermined number of rounds.
[0186] Fig.10 is an example flowchart showing how to obtain a main system behavior model through federated learning based on a federated prediction region according to an embodiment of the present disclosure.
[0187] The process starts at step S101. In step S102, a federated prediction region is constructed. In step S103, an LSTM model is constructed according to the federated prediction region. In step S104, the federated learning client trains a local LSTM model according to local data. In step S105, it is determined whether the aggregation cycle or model convergence has been reached. If the result of the determination in step S105 is no, the process returns to step S104. If the result of the determination in step S105 is yes, the process proceeds to step S106. In step S106, if it is centralized federated learning, the federated learning client uploads the local model to the federated learning server; if it is distributed federated learning, the federated learning client for model aggregation collects the local models of other federated learning clients. In step S107, it is determined whether the local model collection is completed. If the result of the determination in step S107 is no, the process returns to step S106. If the result of the determination in step S107 is yes, the process proceeds to step S108. In step S108, the model aggregation weight is calculated to aggregate the local model, and the global model is used as the new local model. In step S109, it is determined whether the number of federation aggregations has been reached. If the result of the determination in step S109 is no, the process returns to step S104. If the result of the determination in step S109 is yes, the process proceeds to step S110. In step S110, the process ends.
[0188] Aggregating the federated learning model (referred to as quantity-weighted aggregation) using weights calculated based on the number of geographic units occupied by user devices in the secondary system associated with the federated learning client (for example, weights calculated according to formula (1)) can improve the average prediction accuracy of the aggregation model. Figure 11(a) shows the actual received power of the main system and the predicted received power of the main system predicted using the prior art, Figure 11(b) shows the actual received power of the main system and the predicted received power of the main system predicted using quantity-weighted aggregation, and Figure 11(c) shows the error of the predicted received power of the main system predicted using the prior art and the error of the predicted received power of the main system predicted using quantity-weighted aggregation. To demonstrate the performance in terms of prediction accuracy, Figure 11(b) shows the predicted power when the input of the LSTM model is the received power of a PU signal containing 9 minimum cells. Figure 11(a) to Figure 11(c)In FIG11(a) and FIG11(b), 9 areas are involved, namely Area0 to Area8. In FIG11(a) and FIG11(b), the horizontal axis is time, the vertical axis is the predicted received power of the main system and the actual received power of the main system, the unit of power is mW, the solid line is the actual received power (marked with real in the figure), and the dotted line is the predicted received power (marked with prediction in the figure). FIG11(a) shows the prediction effect of the traditional average weighted federated learning algorithm, and FIG11(b) shows the prediction effect of the prediction using the quantity weighted aggregation. It can be seen that FIG11(b) shows a better prediction effect, that is, in FIG11(b), the predicted received power of the main system is closer to the actual received power of the main system. In FIG11(c), the vertical axis is mW, the solid line is the error curve (marked with weight in the figure) of the prediction using the quantity weighted aggregation, and the dotted line is the error curve (marked with avg in the figure) of the traditional average weighted federated learning algorithm. It can be seen that the error of the prediction using the quantity weighted aggregation according to the embodiment of the present disclosure is smaller.
[0189] The present disclosure also provides an electronic device for a spectrum management apparatus according to another embodiment. Fig.12 A functional module block diagram of an electronic device 1200 used in a spectrum management apparatus according to another embodiment of the present disclosure is shown.
[0190] like Fig.12 As shown, the electronic device 1200 includes: an area obtaining unit 1201, which can be configured to obtain a federated prediction area among a plurality of federated prediction areas constructed based on a secondary system; and an assisting processing unit 1203, which can be configured to assist in obtaining a primary system behavior model for predicting the behavior of the primary system using spectrum resources through federated learning. And, it should be understood that Fig.12 The various functional units in the electronic device shown in the figure are only logical modules divided according to the specific functions implemented by them, and are not used to limit the specific implementation method.
[0191] The obtaining federated prediction area unit 1201 and the assisting processing unit 1203 may be implemented by one or more processing circuits, which may be implemented as a chip, for example.
[0192] The electronic device 1200 can be used as a network side device in a wireless communication system, and specifically, for example, can be arranged on the base station side or can be communicatively connected to the base station. Here, it should also be pointed out that the electronic device 1200 can be implemented at the chip level, or it can also be implemented at the device level. For example, the electronic device 1200 can work as a base station itself, and can also include external devices such as a memory, a transceiver (not shown), etc. The memory can be used to store programs and related data information that need to be executed by the electronic device to implement various functions. The transceiver may include one or more communication interfaces to support communication with different devices (e.g., servers, base stations, etc.), and the implementation form of the transceiver is not specifically limited here.
[0193] The wireless communication system according to the present disclosure may be a 5G NR (New Radio) communication system. Further, the wireless communication system according to the present disclosure may include a non-terrestrial network (NTN). Optionally, the wireless communication system according to the present disclosure may also include a terrestrial network (TN). In addition, those skilled in the art may understand that the wireless communication system according to the present disclosure may also be a 4G or 3G communication system.
[0194] The electronic device 1200 according to the embodiment of the present disclosure can assist in more effectively predicting the received power of the primary system in different areas for spectrum resource sensing.
[0195] As an example, the spectrum management device, i.e., the electronic device 1200, can obtain a trained local main system behavior model from a corresponding federated learning client that corresponds to it and participates in federated learning. The corresponding federated learning client preprocesses a data set uploaded from a secondary system associated with the spectrum management device, and trains the local main system behavior model based on the preprocessed data set.
[0196] As an example, the corresponding federated learning client is implemented by NWDAF.
[0197] As an example, the corresponding federated learning client may be one of NWDAF_1 to NWDAF_4 in FIG. 2( a ).
[0198] As an example, the pre-processing includes: when there are multiple primary system signal reception powers corresponding to the same geographical unit in the data set, averaging the multiple primary system signal reception powers, wherein the geographical unit has a predetermined geographical size.
[0199] For the description of the preprocessing, please refer to the description in conjunction with FIG. 2( a ), which will not be repeated here.
[0200] As an example, the corresponding federated learning client sends the trained local main system behavior model to the federated learning server participating in the federated learning, so that the federated learning server aggregates the local main system behavior model to obtain the main system behavior model.
[0201] As an example, the spectrum management device, i.e., the electronic device 1200, can act as a federated learning client participating in federated learning, and the assisting processing unit 1203 can be configured to pre-process the data set uploaded from the secondary system associated with the spectrum management device, and train the local main system behavior model based on the pre-processed data set. In this example, the electronic device 1200 can directly participate in federated learning as a federated learning client. For example, the electronic device 1200 can directly participate in federated learning instead of the corresponding NWDAF (e.g., one of NWDAF_1 to NWDAF_4) in Figure 2(a).
[0202] As an example, the pre-processing includes: when there are multiple primary system signal reception powers corresponding to the same geographical unit in the data set, averaging the multiple primary system signal reception powers, wherein the geographical unit has a predetermined geographical size.
[0203] For the description of the preprocessing, please refer to the description in conjunction with FIG. 2( a ), which will not be repeated here.
[0204] As an example, the assisting processing unit 1203 may be configured to send the trained local main system behavior model to a federated learning server participating in the federated learning, so that the federated learning server may aggregate the local main system behavior model to obtain the main system behavior model.
[0205] In the process of describing the electronic device for the spectrum management device in the above embodiments, it is obvious that some processes or methods are also disclosed. In the following, an overview of these methods is given without repeating some of the details discussed above, but it should be noted that although these methods are disclosed in the process of describing the electronic device for the spectrum management device, these methods do not necessarily adopt the components described or are not necessarily performed by those components. For example, the implementation of the electronic device for the spectrum management device can be partially or completely implemented using hardware and / or firmware, and the methods for the spectrum management device discussed below can be completely implemented by a computer-executable program, although these methods can also adopt the hardware and / or firmware of the electronic device for the spectrum management device.
[0206] Fig.13A flowchart of a method S1300 for a spectrum management device according to an embodiment of the present disclosure is shown. The method S1300 starts at step S1302. In step S1304, for at least one federated prediction area among a plurality of federated prediction areas constructed based on a secondary system, the spectrum management device obtains a primary system behavior model for predicting the behavior of the primary system using spectrum resources through federated learning. The method S1300 ends at step S1306.
[0207] The method may be executed, for example, by the electronic device 100 described above. For specific details, reference may be made to the description of the related processing of the electronic device 100, which will not be repeated here.
[0208] Fig.14 A flowchart of a method S1400 for a spectrum management device according to another embodiment of the present disclosure is shown. The method S1400 starts at step S1402. In step S1404, for one of the multiple federated prediction areas constructed based on the secondary system, a primary system behavior model for predicting the behavior of the primary system using spectrum resources is obtained through federated learning. The method S1400 ends at step S1406.
[0209] The method may be executed, for example, by the electronic device 1200 described above. For specific details, reference may be made to the description of the related processing of the electronic device 1200, which will not be repeated here.
[0210] The technology of the present disclosure can be applied to various products.
[0211] Electronic devices 100 and 1200 may be implemented as various network-side devices. The network-side device may be arranged on the base station side or connected to the base station. The base station may be implemented as any type of evolved Node B (eNB) or gNB (5G base station). eNBs include, for example, macro eNBs and small eNBs. Small eNBs may be eNBs that cover cells smaller than macro cells, such as pico eNBs, micro eNBs, and home (femto) eNBs. Similar situations may also apply to gNBs. Alternatively, the base station may be implemented as any other type of base station, such as a NodeB and a base transceiver station (BTS). The base station may include: a main body (also referred to as a base station device) configured to control wireless communications; and one or more remote radio heads (RRHs) arranged at a different place from the main body. In addition, various types of electronic devices may work as base stations by temporarily or semi-permanently performing base station functions.
[0212] [Application examples for base stations]
[0213] (First application example)
[0214] Fig.15800 is a block diagram showing a first example of a schematic configuration of an eNB or gNB to which the technology of the present disclosure can be applied. Note that the following description takes eNB as an example, but can also be applied to gNB. The eNB 800 includes one or more antennas 810 and a base station device 820. The base station device 820 and each antenna 810 can be connected to each other via an RF cable.
[0215] Each of the antennas 810 includes a single or multiple antenna elements (such as multiple antenna elements included in a multiple-input multiple-output (MIMO) antenna) and is used for the base station device 820 to transmit and receive wireless signals. Fig.15 As shown, the eNB 800 may include multiple antennas 810. For example, the multiple antennas 810 may be compatible with multiple frequency bands used by the eNB 800. Fig.15 An example is shown in which the eNB 800 includes a plurality of antennas 810 , but the eNB 800 may also include a single antenna 810 .
[0216] The base station device 820 includes a controller 821 , a memory 822 , a network interface 823 , and a wireless communication interface 825 .
[0217] The controller 821 may be, for example, a CPU or a DSP, and operates various functions of the higher layers of the base station device 820. For example, the controller 821 generates a data packet based on the data in the signal processed by the wireless communication interface 825, and transmits the generated packet via the network interface 823. The controller 821 may bundle data from a plurality of baseband processors to generate a bundled packet, and transmit the generated bundled packet. The controller 821 may have a logical function to perform the following control: the control may be such as radio resource control, radio bearer control, mobility management, admission control, and scheduling. The control may be performed in conjunction with a nearby eNB or core network node. The memory 822 includes a RAM and a ROM, and stores programs executed by the controller 821 and various types of control data (such as a terminal list, transmission power data, and scheduling data).
[0218] The network interface 823 is a communication interface for connecting the base station device 820 to the core network 824. The controller 821 can communicate with the core network node or another eNB via the network interface 823. In this case, the eNB 800 and the core network node or other eNBs can be connected to each other through a logical interface (such as an S1 interface and an X2 interface). The network interface 823 can also be a wired communication interface or a wireless communication interface for a wireless backhaul line. If the network interface 823 is a wireless communication interface, the network interface 823 can use a higher frequency band for wireless communication compared to the frequency band used by the wireless communication interface 825.
[0219] The wireless communication interface 825 supports any cellular communication scheme (such as long term evolution (LTE) and LTE-Advanced), and provides a wireless connection to a terminal located in a cell of the eNB 800 via the antenna 810. The wireless communication interface 825 may generally include, for example, a baseband (BB) processor 826 and an RF circuit 827. The BB processor 826 may perform, for example, encoding / decoding, modulation / demodulation, and multiplexing / demultiplexing, and perform various types of signal processing of layers (e.g., layer 1, medium access control (MAC), radio link control (RLC), and packet data convergence protocol (PDCP)). Instead of the controller 821, the BB processor 826 may have a part or all of the above-mentioned logical functions. The BB processor 826 may be a memory storing a communication control program, or a module including a processor configured to execute a program and related circuits. Updating the program may change the function of the BB processor 826. The module may be a card or a blade inserted into a slot of the base station device 820. Alternatively, the module may also be a chip mounted on a card or a blade. Meanwhile, the RF circuit 827 may include, for example, a mixer, a filter, and an amplifier, and transmits and receives a wireless signal via the antenna 810 .
[0220] like Fig.15 As shown, the wireless communication interface 825 may include multiple BB processors 826. For example, the multiple BB processors 826 may be compatible with multiple frequency bands used by the eNB 800. Fig.15 As shown, the wireless communication interface 825 may include multiple RF circuits 827. For example, the multiple RF circuits 827 may be compatible with multiple antenna elements. Fig.15 An example is shown in which the wireless communication interface 825 includes a plurality of BB processors 826 and a plurality of RF circuits 827 , but the wireless communication interface 825 may also include a single BB processor 826 or a single RF circuit 827 .
[0221] exist Fig.15 In the illustrated eNB 800, when the electronic devices 100 and 1200 are implemented as base stations, their transceivers may be implemented by the wireless communication interface 825. At least part of the functions may also be implemented by the controller 821. For example, the controller 821 may more effectively predict the received power of the primary system in different areas for spectrum resource sensing by executing the functions of the units in the electronic devices 100 and 1200.
[0222] (Second application example)
[0223] Fig.16830 is a block diagram showing a second example of a schematic configuration of an eNB or gNB to which the technology of the present disclosure can be applied. Note that similarly, the following description takes the eNB as an example, but can also be applied to the gNB. The eNB 830 includes one or more antennas 840, a base station device 850, and an RRH 860. The RRH 860 and each antenna 840 can be connected to each other via an RF cable. The base station device 850 and the RRH 860 can be connected to each other via a high-speed line such as an optical fiber cable.
[0224] Each of the antennas 840 includes a single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna) and is used for the RRH 860 to transmit and receive wireless signals. Fig.16 As shown, the eNB 830 may include multiple antennas 840. For example, the multiple antennas 840 may be compatible with multiple frequency bands used by the eNB 830. Fig.16 An example is shown in which the eNB 830 includes a plurality of antennas 840 , but the eNB 830 may also include a single antenna 840 .
[0225] The base station device 850 includes a controller 851, a memory 852, a network interface 853, a wireless communication interface 855, and a connection interface 857. The controller 851, the memory 852, and the network interface 853 are similar to the reference Fig.15 The controller 821, memory 822 and network interface 823 described are the same.
[0226] The wireless communication interface 855 supports any cellular communication scheme (such as LTE and LTE-Advanced), and provides wireless communication to a terminal located in a sector corresponding to the RRH 860 via the RRH 860 and the antenna 840. The wireless communication interface 855 may generally include, for example, a BB processor 856. In addition to the BB processor 856 being connected to the RF circuit 864 of the RRH 860 via the connection interface 857, the BB processor 856 is connected to the reference RF circuit 864 of the RRH 860. Fig.15 The same as the BB processor 826 described above. Fig.16 As shown, the wireless communication interface 855 may include multiple BB processors 856. For example, the multiple BB processors 856 may be compatible with multiple frequency bands used by the eNB 830. Fig.16 An example is shown in which the wireless communication interface 855 includes a plurality of BB processors 856 , but the wireless communication interface 855 may also include a single BB processor 856 .
[0227] The connection interface 857 is an interface for connecting the base station device 850 (wireless communication interface 855) to the RRH 860. The connection interface 857 may also be a communication module for connecting the base station device 850 (wireless communication interface 855) to the RRH 860 for communication in the above-mentioned high-speed line.
[0228] The RRH 860 includes a connection interface 861 and a wireless communication interface 863 .
[0229] The connection interface 861 is an interface for connecting the RRH 860 (wireless communication interface 863) to the base station device 850. The connection interface 861 may also be a communication module for communication in the above-mentioned high-speed line.
[0230] The wireless communication interface 863 transmits and receives wireless signals via the antenna 840. The wireless communication interface 863 may generally include, for example, an RF circuit 864. The RF circuit 864 may include, for example, a mixer, a filter, and an amplifier, and transmits and receives wireless signals via the antenna 840. Fig.16 As shown, the wireless communication interface 863 may include multiple RF circuits 864. For example, the multiple RF circuits 864 may support multiple antenna elements. Fig.16 An example is shown in which the wireless communication interface 863 includes a plurality of RF circuits 864 , but the wireless communication interface 863 may also include a single RF circuit 864 .
[0231] exist Fig.16 In the illustrated eNB 830, when the electronic devices 100 and 1200 are implemented as base stations, their transceivers may be implemented by the wireless communication interface 855. At least part of the functions may also be implemented by the controller 851. For example, the controller 851 may more effectively predict the received power of the primary system in different areas for spectrum resource sensing by executing the functions of the units in the electronic devices 100 and 1200.
[0232] [Application examples on user devices]
[0233] (First application example)
[0234] Fig.17 900 is a block diagram showing an example of a schematic configuration of a smartphone 900 to which the technology of the present disclosure can be applied. The smartphone 900 includes a processor 901, a memory 902, a storage device 903, an external connection interface 904, a camera device 906, a sensor 907, a microphone 908, an input device 909, a display device 910, a speaker 911, a wireless communication interface 912, one or more antenna switches 915, one or more antennas 916, a bus 917, a battery 918, and an auxiliary controller 919.
[0235] The processor 901 may be, for example, a CPU or a system on chip (SoC), and controls the functions of the application layer and other layers of the smartphone 900. The memory 902 includes a RAM and a ROM, and stores data and programs executed by the processor 901. The storage device 903 may include a storage medium such as a semiconductor memory and a hard disk. The external connection interface 904 is an interface for connecting an external device such as a memory card and a universal serial bus (USB) device to the smartphone 900.
[0236] The camera 906 includes an image sensor such as a charge coupled device (CCD) and a complementary metal oxide semiconductor (CMOS), and generates a captured image. The sensor 907 may include a group of sensors such as a measurement sensor, a gyro sensor, a geomagnetic sensor, and an acceleration sensor. The microphone 908 converts the sound input to the smart phone 900 into an audio signal. The input device 909 includes, for example, a touch sensor, a keypad, a keyboard, a button, or a switch configured to detect a touch on the screen of the display device 910, and receives an operation or information input from a user. The display device 910 includes a screen such as a liquid crystal display (LCD) and an organic light emitting diode (OLED) display, and displays an output image of the smart phone 900. The speaker 911 converts an audio signal output from the smart phone 900 into sound.
[0237] The wireless communication interface 912 supports any cellular communication scheme (such as LTE and LTE-Advanced), and performs wireless communication. The wireless communication interface 912 may generally include, for example, a BB processor 913 and an RF circuit 914. The BB processor 913 may perform, for example, encoding / decoding, modulation / demodulation, and multiplexing / demultiplexing, and perform various types of signal processing for wireless communication. Meanwhile, the RF circuit 914 may include, for example, a mixer, a filter, and an amplifier, and transmit and receive wireless signals via an antenna 916. Note that although the figure shows a situation where one RF link is connected to one antenna, this is only schematic, and also includes a situation where one RF link is connected to multiple antennas through multiple phase shifters. The wireless communication interface 912 may be a chip module on which the BB processor 913 and the RF circuit 914 are integrated. As shown in FIG. Fig.17 As shown, the wireless communication interface 912 may include multiple BB processors 913 and multiple RF circuits 914. Fig.17 An example is shown in which the wireless communication interface 912 includes a plurality of BB processors 913 and a plurality of RF circuits 914 , but the wireless communication interface 912 may also include a single BB processor 913 or a single RF circuit 914 .
[0238] In addition, in addition to the cellular communication scheme, the wireless communication interface 912 can support other types of wireless communication schemes, such as short-range wireless communication schemes, near field communication schemes, and wireless local area network (LAN) schemes. In this case, the wireless communication interface 912 can include a BB processor 913 and an RF circuit 914 for each wireless communication scheme.
[0239] Each of the antenna switches 915 switches a connection destination of the antenna 916 between a plurality of circuits (eg, circuits for different wireless communication schemes) included in the wireless communication interface 912 .
[0240] Each of the antennas 916 includes a single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna) and is used for the wireless communication interface 912 to transmit and receive wireless signals. Fig.17 As shown, the smart phone 900 may include multiple antennas 916. Fig.17 An example is shown in which the smartphone 900 includes a plurality of antennas 916 , but the smartphone 900 may also include a single antenna 916 .
[0241] In addition, the smartphone 900 may include an antenna 916 for each wireless communication scheme. In this case, the antenna switch 915 may be omitted from the configuration of the smartphone 900.
[0242] The bus 917 connects the processor 901, the memory 902, the storage device 903, the external connection interface 904, the camera 906, the sensor 907, the microphone 908, the input device 909, the display device 910, the speaker 911, the wireless communication interface 912, and the auxiliary controller 919 to each other. Fig.17 The various blocks of the smartphone 900 shown are supplied with power, the feed lines being partially shown as dashed lines in the figure. The auxiliary controller 919 operates the minimum necessary functions of the smartphone 900, for example in a sleep mode.
[0243] exist Fig.17 In the illustrated smartphone 900 , the transceiver may be implemented by the wireless communication interface 912 . At least part of the functionality may also be implemented by the processor 901 or the auxiliary controller 919 .
[0244] (Second application example)
[0245] Fig.189 is a block diagram showing an example of a schematic configuration of a car navigation device 920 to which the technology of the present disclosure can be applied. The car navigation device 920 includes a processor 921, a memory 922, a global positioning system (GPS) module 924, a sensor 925, a data interface 926, a content player 927, a storage medium interface 928, an input device 929, a display device 930, a speaker 931, a wireless communication interface 933, one or more antenna switches 936, one or more antennas 937, and a battery 938.
[0246] The processor 921 may be, for example, a CPU or an SoC, and controls a navigation function and another function of the car navigation device 920. The memory 922 includes a RAM and a ROM, and stores data and a program executed by the processor 921.
[0247] The GPS module 924 measures the position (such as latitude, longitude and altitude) of the car navigation device 920 using GPS signals received from GPS satellites. The sensor 925 may include a group of sensors such as a gyro sensor, a geomagnetic sensor and an air pressure sensor. The data interface 926 is connected to, for example, the vehicle network 941 via an unshown terminal and acquires data (such as vehicle speed data) generated by the vehicle.
[0248] The content player 927 reproduces content stored in a storage medium such as a CD and a DVD, which is inserted into the storage medium interface 928. The input device 929 includes, for example, a touch sensor, a button, or a switch configured to detect a touch on the screen of the display device 930, and receives an operation or information input from a user. The display device 930 includes a screen such as an LCD or an OLED display, and displays an image of a navigation function or reproduced content. The speaker 931 outputs a sound of a navigation function or reproduced content.
[0249] The wireless communication interface 933 supports any cellular communication scheme (such as LTE and LTE-Advanced), and performs wireless communication. The wireless communication interface 933 may generally include, for example, a BB processor 934 and an RF circuit 935. The BB processor 934 may perform, for example, encoding / decoding, modulation / demodulation, and multiplexing / demultiplexing, and perform various types of signal processing for wireless communication. Meanwhile, the RF circuit 935 may include, for example, a mixer, a filter, and an amplifier, and transmit and receive wireless signals via an antenna 937. The wireless communication interface 933 may also be a chip module on which the BB processor 934 and the RF circuit 935 are integrated. Fig.18 As shown, the wireless communication interface 933 may include multiple BB processors 934 and multiple RF circuits 935. Fig.18An example is shown in which the wireless communication interface 933 includes a plurality of BB processors 934 and a plurality of RF circuits 935 , but the wireless communication interface 933 may also include a single BB processor 934 or a single RF circuit 935 .
[0250] In addition, in addition to the cellular communication scheme, the wireless communication interface 933 can support other types of wireless communication schemes, such as short-range wireless communication schemes, near field communication schemes, and wireless LAN schemes. In this case, for each wireless communication scheme, the wireless communication interface 933 can include a BB processor 934 and an RF circuit 935.
[0251] Each of the antenna switches 936 switches a connection destination of the antenna 937 between a plurality of circuits included in the wireless communication interface 933 , such as circuits for different wireless communication schemes.
[0252] Each of the antennas 937 includes a single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna), and is used for the wireless communication interface 933 to transmit and receive wireless signals. Fig.18 As shown, the car navigation device 920 may include multiple antennas 937. Fig.18 An example is shown in which the car navigation device 920 includes a plurality of antennas 937 , but the car navigation device 920 may also include a single antenna 937 .
[0253] Furthermore, the car navigation device 920 may include an antenna 937 for each wireless communication scheme. In this case, the antenna switch 936 may be omitted from the configuration of the car navigation device 920.
[0254] Battery 938 is fed via a feed line to Fig.18 The respective blocks of the illustrated car navigation device 920 are supplied with electric power, and feed lines are partially illustrated as dotted lines in the figure. The battery 938 accumulates electric power supplied from the vehicle.
[0255] exist Fig.18 In the illustrated car navigation device 920 , the transceiver may be implemented by the wireless communication interface 933 . At least part of the functions may also be implemented by the processor 921 .
[0256] The technology of the present disclosure may also be implemented as an in-vehicle system (or vehicle) 940 including a car navigation device 920, an in-vehicle network 941, and one or more blocks in a vehicle module 942. The vehicle module 942 generates vehicle data (such as vehicle speed, engine speed, and fault information), and outputs the generated data to the in-vehicle network 941.
[0257] The basic principles of the present invention are described above in conjunction with specific embodiments. However, it should be pointed out that for those skilled in the art, it is understandable that all or any steps or components of the method and device of the present invention can be implemented in any computing device (including a processor, storage medium, etc.) or a network of computing devices in the form of hardware, firmware, software, or a combination thereof. This can be achieved by those skilled in the art using their basic circuit design knowledge or basic programming skills after reading the description of the present invention.
[0258] Furthermore, the present invention also provides a program product storing machine-readable instruction codes. When the instruction codes are read and executed by a machine, the method according to the embodiment of the present invention can be executed.
[0259] Accordingly, the storage medium for carrying the program product storing the machine-readable instruction code is also included in the disclosure of the present invention. The storage medium includes but is not limited to a floppy disk, an optical disk, a magneto-optical disk, a memory card, a memory stick, and the like.
[0260] When the present invention is implemented by software or firmware, the data is transferred from a storage medium or a network to a computer (eg, Fig.19 The general-purpose computer 1900 shown in the figure is installed with the programs constituting the software, and when the various programs are installed, the computer can execute various functions, etc.
[0261] exist Fig.19 In the embodiment of the present invention, a central processing unit (CPU) 1901 performs various processes according to a program stored in a read-only memory (ROM) 1902 or a program loaded from a storage section 1908 to a random access memory (RAM) 1903. In the RAM 1903, data required when the CPU 1901 performs various processes and the like is also stored as needed. The CPU 1901, the ROM 1902, and the RAM 1903 are connected to each other via a bus 1904. An input / output interface 1905 is also connected to the bus 1904.
[0262] The following components are connected to the input / output interface 1905: an input section 1906 (including a keyboard, a mouse, etc.), an output section 1907 (including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.), a storage section 1908 (including a hard disk, etc.), a communication section 1909 (including a network interface card such as a LAN card, a modem, etc.). The communication section 1909 performs communication processing via a network such as the Internet. A drive 1910 may also be connected to the input / output interface 1905 as needed. A removable medium 1911 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is installed on the drive 1910 as needed, so that a computer program read therefrom is installed in the storage section 1908 as needed.
[0263] In the case where the above-described series of processing is realized by software, a program constituting the software is installed from a network such as the Internet or a storage medium such as the removable medium 1911 .
[0264] It should be understood by those skilled in the art that such storage media is not limited to Fig.19 The removable medium 1911 shown has a program stored therein and is distributed separately from the device to provide the program to the user. Examples of the removable medium 1911 include magnetic disks (including floppy disks (registered trademark)), optical disks (including compact disk read only memory (CD-ROM) and digital versatile disks (DVD)), magneto-optical disks (including minidiscs (MD) (registered trademark)), and semiconductor memories. Alternatively, the storage medium may be a ROM 1902, a hard disk included in the storage portion 1908, or the like, in which the program is stored and distributed to the user together with the device containing them.
[0265] It should also be noted that in the apparatus, method and system of the present invention, each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent solutions of the present invention. In addition, the steps of performing the above series of processes can naturally be performed in chronological order according to the order of description, but do not necessarily need to be performed in chronological order. Some steps can be performed in parallel or independently of each other.
[0266] Finally, it should be noted that the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In addition, in the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, article or device that includes the elements.
[0267] Although the embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, it should be understood that the above-described embodiments are only used to illustrate the present invention and do not constitute a limitation of the present invention. For those skilled in the art, various modifications and changes can be made to the above-mentioned embodiments without departing from the essence and scope of the present invention. Therefore, the scope of the present invention is limited only by the attached claims and their equivalent meanings.
[0268] The present technology can also be implemented as follows.
[0269] Solution 1. An electronic device for a spectrum management device, comprising:
[0270] The processing circuit is configured to:
[0271] For at least one of the multiple federated prediction areas constructed based on the secondary system, a primary system behavior model for predicting the behavior of the primary system in using spectrum resources is obtained through federated learning.
[0272] Solution 2. An electronic device according to Solution 1, wherein the plurality of federated prediction areas are constructed based on the geographical location of the subsystem, and the size of the federated prediction area is reflected by the number of geographical units having a predetermined geographical size.
[0273] Solution 3. An electronic device according to Solution 2, wherein the number and locations of the plurality of federated prediction areas are determined by a k-means algorithm and a density-based spatial clustering algorithm DBSCAN for noisy applications.
[0274] Solution 4. The electronic device according to Solution 3, wherein:
[0275] The k-means algorithm is used to cluster the base stations in the secondary system, delete the cluster that does not contain any base station when the position of the cluster center does not change, and set the position of the specific base station in the cluster as the new cluster center when the distance between the specific base station and the center of the cluster to which it belongs exceeds a first predetermined distance threshold.
[0276] The DBSCAN algorithm uses the positions of the cluster centers determined by the k-means algorithm as input to obtain new cluster centers, and merges the cluster centers whose distances are less than a second predetermined distance threshold among the new cluster centers, and
[0277] The k-means algorithm and the DBSCAN algorithm are repeated until the number of cluster centers no longer changes, wherein the cluster centers serve as the centers of the federated prediction regions and the number of the cluster centers corresponds to the number of the plurality of federated prediction regions.
[0278] Solution 5. An electronic device according to any one of Solutions 2 to 4, wherein the lower limit value of the size of the multiple federal prediction areas is determined based on the distance between the center of the federal prediction area and the secondary system, and each federal prediction area is updated and determined based on the accuracy and / or energy consumption when the main system behavior model converges.
[0279] Solution 6. An electronic device according to any one of Solutions 1 to 5, wherein the main system behavior model is a multi-input multi-output long short-term memory neural network model.
[0280] Option 7. An electronic device according to Option 6, wherein the dimensions of the input and output of the multi-input multi-output long short-term memory neural network model are determined by the size of the federated prediction area.
[0281] Solution 8. An electronic device according to Solution 7, wherein, for each federal prediction area, the input of the multi-input multi-output long-term and short-term neural network model is the main system signal reception power information of all geographical units included in the federal prediction area in a predetermined number of historical time slots, and the output is the predicted value of the main system signal reception power of each geographical unit in the next time slot.
[0282] Solution 9. An electronic device according to any one of Solutions 1 to 8, wherein the main system behavior model is obtained by aggregating local main system behavior models obtained from federated learning clients participating in the federated learning.
[0283] Solution 10. An electronic device according to Solution 9, wherein the local main system behavior model corresponding to the federated learning client is aggregated by calculating weights based on the number of geographical units occupied by user devices in the sub-system associated with the federated learning client.
[0284] Solution 11. An electronic device according to Solution 9, wherein the local main system behavior model corresponding to the federated learning client is aggregated by calculating weights based on the distribution status of user devices in the sub-system associated with the federated learning client.
[0285] Solution 12. An electronic device according to Solution 9, wherein the local main system behavior model is aggregated by calculating a weight based on the model accuracy of the local main system behavior model corresponding to the federated learning client.
[0286] Solution 13. An electronic device according to Solution 12, wherein the model accuracy of the local main system behavior model increases as the amount of user data involved in the local main system behavior model increases, and decreases as the loss of the local main system behavior model increases.
[0287] Solution 14. The electronic device according to any one of Solutions 9 to 13, wherein as the loss of the local main system behavior model decreases monotonically, the period of the aggregation also decreases monotonically.
[0288] Solution 15. The electronic device according to Solution 14, wherein if the loss of the local main system behavior model is not monotonically decreasing, the period of the aggregation is reduced according to a predetermined ratio.
[0289] Scheme 16. An electronic device according to any one of Schemes 1 to 15, wherein the spectrum management device acts as a federated learning server, and the processing circuit is configured to use as the main system behavior model a final global model obtained by aggregating local main system behavior models trained by one or more other spectrum management devices that act as respective federated learning clients participating in the federated learning.
[0290] Option 17. An electronic device according to Option 16, wherein the processing circuit is configured to send a federated learning request to other spectrum management devices within a federated prediction area associated with it, and receive a federated learning response from other spectrum management devices that agree to participate in the federated learning.
[0291] Solution 18. The electronic device according to Solution 17, wherein the federated learning response includes the number of geographical units occupied by user devices in the subsystem associated with the federated learning client.
[0292] Solution 19. The electronic device according to any one of Solution 16 to Solution 18, wherein the processing circuit is configured to:
[0293] In each round of training of the federated learning, the local main system behavior models trained by each federated learning client are aggregated based on the weights to obtain a global model, and the global model is sent to each federated learning client, so that each federated learning client performs the next round of training based on the global model; and
[0294] Until a predetermined condition is met, the federated learning is stopped, and the final global model obtained is used as the main system behavior model.
[0295] Solution 20. An electronic device according to Solution 19, wherein the predetermined condition includes that the final global model reaches a predetermined accuracy and / or has been trained for a predetermined number of rounds.
[0296] Scheme 21. An electronic device according to any one of Schemes 1 to 15, wherein the spectrum management device uses the final global model obtained from the federated learning server as the main system behavior model, wherein the federated learning server aggregates local main system behavior models trained by each federated learning client participating in the federated learning to obtain the final global model.
[0297] Solution 22. An electronic device according to Solution 21, wherein the federated learning server sends a federated learning request to other devices within a federated prediction area associated with it, and receives a federated learning response from other devices that agree to participate in the federated learning.
[0298] Solution 23. An electronic device according to Solution 22, wherein the federated learning response includes the number of geographical units occupied by user devices in the subsystem associated with the federated learning client.
[0299] Solution 24. The electronic device according to any one of Solution 21 to Solution 23, wherein the federated learning server:
[0300] In each round of training of the federated learning, the local main system behavior models trained by each federated learning client are aggregated based on the weights to obtain a global model, and the global model is sent to each federated learning client, so that each federated learning client performs the next round of training based on the global model; and
[0301] Until a predetermined condition is met, the federated learning is stopped to obtain the final global model.
[0302] Solution 25. An electronic device according to Solution 24, wherein the predetermined condition includes that the final global model reaches a predetermined accuracy and / or has been trained for a predetermined number of rounds.
[0303] Solution 26. An electronic device according to any one of Solutions 21 to 25, wherein the federated learning server is implemented by a network data analysis function NWDAF.
[0304] Scheme 27. An electronic device according to any one of Schemes 1 to 15, wherein the spectrum management device acts as a federated learning client, and the processing circuit is configured to perform federated learning together with one or more other spectrum management devices acting as federated learning clients, and use a global model obtained by aggregating local main system behavior models trained by each federated learning client as the main system behavior model.
[0305] Solution 28. An electronic device according to Solution 27, wherein the processing circuit is configured to send the number of geographical units occupied by user equipment in the subsystem associated with the spectrum management device to other federated learning clients.
[0306] Solution 29. An electronic device according to Solution 27 or 28, wherein the processing circuit is configured to, in each round of training of the federated learning, aggregate the local main system behavior model trained by the spectrum management device and the local main system behavior model received from other federated learning clients through broadcasting based on weights to obtain a global model, and use the global model as a new local main system behavior model for the next round of training; and
[0307] Until a predetermined condition is met, the federated learning is stopped, and the final global model obtained is used as the main system behavior model.
[0308] Solution 30. The electronic device according to Solution 27 or 28, wherein the spectrum management device is selected to perform the aggregation when the following predetermined conditions are met:
[0309] Based on the channel information between the spectrum management device and the one or more other spectrum management devices, calculating the sum of transmission energy consumptions of transmission paths between the spectrum management device and the one or more other spectrum management devices when the spectrum management device is used as an aggregation device for aggregating local primary user behavior models within a federated prediction area associated with the spectrum management device,
[0310] The sum of transmission energy consumption when the spectrum management device serves as the aggregation device is less than the sum of transmission energy consumption when any other spectrum management device of the one or more other spectrum management devices serves as the aggregation device.
[0311] Embodiment 31. The electronic device according to Embodiment 30, wherein the processing circuit is configured to exchange the channel information with the one or more other spectrum management devices through broadcasting.
[0312] Solution 32. An electronic device according to Solution 30 or 31, wherein the transmission energy consumption of the transmission path is calculated based on the transmission rate and transmission power of the transmission path.
[0313] Solution 33. An electronic device according to any one of Solutions 27 to 33, wherein the spectrum management device is implemented by a base station in a subsystem or an edge server associated with the subsystem.
[0314] Solution 34. An electronic device for a spectrum management device, comprising:
[0315] The processing circuit is configured to:
[0316] For one of the plurality of federated prediction areas constructed based on the secondary system, a primary system behavior model for predicting the behavior of the primary system in using spectrum resources is obtained through federated learning.
[0317] Solution 35. An electronic device according to Solution 34, wherein:
[0318] The spectrum management device serves as a federated learning client participating in the federated learning,
[0319] The processing circuit is configured to pre-process a data set uploaded from a secondary system associated with the spectrum management device, and train a local primary system behavior model based on the pre-processed data set.
[0320] Scheme 36. An electronic device according to Scheme 35, wherein the preprocessing includes: when there are multiple main system signal reception powers corresponding to the same geographic unit in the data set, averaging the multiple main system signal reception powers, wherein the geographic unit has a predetermined geographic size.
[0321] Scheme 37. An electronic device according to Scheme 35 or 36, wherein the processing circuit is configured to send the trained local main system behavior model to a federated learning server participating in the federated learning, so that the federated learning server can aggregate the local main system behavior model to obtain the main system behavior model.
[0322] Solution 38. An electronic device according to Solution 34, wherein:
[0323] The spectrum management device obtains a trained local main system behavior model from a corresponding federated learning client that participates in the federated learning.
[0324] The corresponding federated learning client preprocesses a data set uploaded from a secondary system associated with the spectrum management device, and trains a local primary system behavior model based on the preprocessed data set.
[0325] Scheme 39. An electronic device according to Scheme 38, wherein the preprocessing includes: when there are multiple main system signal reception powers corresponding to the same geographic unit in the data set, averaging the multiple main system signal reception powers, wherein the geographic unit has a predetermined geographic size.
[0326] Solution 40. An electronic device according to Solution 38 or 39, wherein the corresponding federated learning client sends the trained local main system behavior model to a federated learning server participating in the federated learning, so that the federated learning server aggregates the local main system behavior model to obtain the main system behavior model.
[0327] Solution 41. An electronic device according to any one of Solutions 38 to 40, wherein the corresponding federated learning client is implemented by a network data analysis function NWDAF.
[0328] Scheme 42. A method for a spectrum management device, comprising: for at least one federal prediction area among multiple federal prediction areas constructed based on a secondary system, the spectrum management device obtains a main system behavior model for predicting the behavior of the main system in using spectrum resources through federated learning.
[0329] Scheme 43. A method for a spectrum management device, comprising: for one federal prediction area among multiple federal prediction areas constructed based on a secondary system, assisting in obtaining a main system behavior model for predicting the behavior of a main system in using spectrum resources through federated learning.
[0330] Scheme 44. A computer-readable storage medium having computer-executable instructions stored thereon. When the computer-executable instructions are executed by a processor, the processor executes the method for a spectrum management device according to any one of Schemes 42 to 43.
Claims
1. An electronic device for a spectrum management device, include: The processing circuit is configured to: For at least one of the multiple federated prediction areas constructed based on the secondary system, a primary system behavior model for predicting the behavior of the primary system in using spectrum resources is obtained through federated learning.
2. The electronic device according to claim 1, in, The plurality of federated prediction areas are constructed based on the geographic location of the subsystem, and the size of the federated prediction area is represented by the number of geographic units having a predetermined geographic size.
3. The electronic device according to claim 1, in, The main system behavior model is a multi-input multi-output long short-term memory neural network model.
4. The electronic device according to claim 1, in, The spectrum management device acts as a federated learning server, and the processing circuit is configured to use a final global model obtained by aggregating local main system behavior models trained by one or more other spectrum management devices that act as federated learning clients participating in the federated learning as the main system behavior model.
5. The electronic device according to claim 1, in, The spectrum management device uses the final global model obtained from the federated learning server as the main system behavior model, wherein the federated learning server aggregates local main system behavior models trained by various federated learning clients participating in the federated learning to obtain the final global model.
6. The electronic device according to claim 1, in, The spectrum management device acts as a federated learning client, and the processing circuit is configured to perform federated learning together with one or more other spectrum management devices acting as federated learning clients, and use the global model obtained by aggregating the local main system behavior models trained by each federated learning client as the main system behavior model.
7. An electronic device for a spectrum management device, include: The processing circuit is configured to: For one of the plurality of federated prediction areas constructed based on the secondary system, a primary system behavior model for predicting the behavior of the primary system in using spectrum resources is obtained through federated learning.
8. A method for a spectrum management device, include: For at least one of the multiple federated prediction areas constructed based on the secondary system, the spectrum management device obtains a primary system behavior model for predicting the behavior of the primary system using spectrum resources through federated learning.
9. A method for a spectrum management device, include: For one of the plurality of federated prediction areas constructed based on the secondary system, a primary system behavior model for predicting the behavior of the primary system in using spectrum resources is obtained through federated learning.
10. A computer-readable storage medium having computer-executable instructions stored thereon, and when the computer-executable instructions are executed by a processor, the processor is enabled to execute the method for a spectrum management device according to any one of claims 8 to 9.
