Federated learning system and federated learning method

Abstract

The purpose of this invention is to suppress the communication volume between the base station and the host device in the RIC processing of communication between the base station and the terminal. The federated learning system includes multiple base stations and a host device capable of communicating with each of the multiple base stations. Each base station includes: a learning unit that learns communication information related to communication with the terminal and generates a local learning model related to communication through a real-time RIC (Radio Access Network Intelligent Controller); and a transmitting unit that transmits information related to the local learning model to the host device. The host device includes: a receiving unit that receives information related to the local learning model; a generating unit that generates an ensemble learning model based on the local learning models of each of the multiple base stations through a non-real-time RIC; and a feedback unit that feeds back information related to the ensemble learning model generated by the generating unit to the base station.

Description

Technical Field

[0001] This invention relates to a federated learning system and a federated learning method that utilizes base stations for learning. Background Technology

[0002] In recent years, research has been conducted on the application of RICs (Radio Access Network Intelligent Controllers) in communication between base stations and terminals. In previous RICs, learning typically involved collecting large amounts of information from the base station and performing the process non-real-time by a central device. Patent Document 1 discloses an invention that utilizes a base station for learning and selects participants for federated learning. Existing technical documents Patent documents

[0003] Patent Document 1: Japanese Patent Publication No. 2023-550806 Summary of the Invention The problem that the invention aims to solve

[0004] However, when learning is performed by a central device, during real-time learning and control, the large amount of data sent from the base station to the central device for learning causes increased traffic between the base station and the central device, making communication strained. As a result, there is a problem of insufficient resources for learning in the central device.

[0005] Therefore, the present invention was made in view of the above-mentioned problems, and its object is to provide a federated learning system and a federated learning method that can perform learning without straining communication between the base station and the central device. Methods for solving problems

[0006] To address the aforementioned issues, one aspect of the federated learning system of the present invention includes multiple base stations and a host device capable of communicating with each of the multiple base stations. Each base station comprises: a learning unit that learns communication information related to communication with terminals and generates a local learning model related to communication via a real-time RIC (Radio Access Network Intelligent Controller); and a transmitting unit that transmits information related to the local learning model to the host device. The host device comprises: a receiving unit that receives information related to the local learning model; a generating unit that generates an ensemble learning model based on the local learning models of each of the multiple base stations via a non-real-time RIC; and a feedback unit that feeds back information related to the ensemble learning model generated by the generating unit to the base stations.

[0007] Furthermore, in the aforementioned federated learning system, the information related to the local learning model may be a portion of the parameter information of the pre-specified local learning model. The generation of the ensemble learning model may be achieved by averaging the parameter information sent from the multiple base stations. The feedback unit may feed back the averaged parameter information to the base station. The learning unit may replace the parameters in the local learning model corresponding to the parameter information with the averaged parameters.

[0008] Furthermore, in the aforementioned federated learning system, the generation unit can reduce the frequency of generating ensemble learning models through the non-real-time RIC over time.

[0009] In addition, to address the aforementioned issues, one aspect of the federated learning method of the present invention is a federated learning system comprising multiple base stations and a host device capable of communicating with each of the multiple base stations. The federated learning method performs the following steps: a learning step, in which the base station learns communication information related to communication with the terminal and generates a local learning model related to the communication via a real-time RIC (Radio Access Network Intelligent Controller); a sending step, in which the base station sends information related to the local learning model to the host device; a receiving step, in which the host device receives information related to the local learning model; a generation step, in which the host device generates an ensemble learning model based on the respective local learning models of the multiple base stations via a non-real-time RIC; and a feedback step, in which information related to the ensemble learning model generated in the generation step is fed back to the base stations. The effects of the invention

[0010] In the federated learning system of the present invention, each base station learns in real time based on information obtained in communication with the terminal, and only sends the information related to the local learning model obtained through learning from the base station to the host device, thereby eliminating the need to send a large amount of terminal information, thus suppressing the traffic congestion between the base station and the host device. Attached Figure Description

[0011] Figure 1 This is a system diagram representing an example of the structure of a communication system. Figure 2 This is a block diagram representing a structural example of a base station. Figure 3 This is a block diagram representing a structural example of a service management device. Figure 4 This is a flowchart illustrating an example of a base station's actions. Figure 5 This is a flowchart illustrating an example of the actions of a service management device. Detailed Implementation

[0012] The base station of the present invention will now be described with reference to the accompanying drawings.

[0013] <Summary> Figure 1 This is a schematic diagram representing the outline of Federated Learning System 1. Federated Learning System 1 can be a communication system that relays communication between terminals 300 (300a~300d) via base stations 100 (100a, 100b). Base stations 100 can be managed by a service management device 200, which acts as a higher-level device. Federated Learning System 1 can be a telephone communication network formed by a virtualized wireless communication network. The service management device 200 can be a CU (Central Unit) in 5G communication, and base stations 100 can be DU (Distributed Unit) or RU (Radio Unit) in 5G communication.

[0014] In this embodiment, each base station 100 performs a predetermined learning process in real time (allowing for some delay) based on the received signals received from the terminal 300 communicating with it. The learning content in the predetermined learning can be determined by the operator or other personnel in the federated learning system 1, or it can be arbitrary. As an example, the predetermined learning can learn the relationship between the signal received from the terminal 300 (for example, it can be an SRS (Sounding Reference Signal), but is not limited to this) and the channel information of the terminal 300, but is not limited to this. In the case of this example learning, as long as there is a received signal from the terminal 300, the channel information applicable to the terminal 300 can be inferred.

[0015] As for the learning model generated through such learning, it is preferable to generate a model with the highest possible generality based on more information. Typically, it is considered that each base station 100 collects information from the terminal 300 and uploads it to the service management device 200 for accumulation, and the service management device 200 learns from it. However, in this method, since a large amount of data is sent from the base station 100 to the service management device 200, the traffic in the service management device 200 increases, and communication may become strained.

[0016] Therefore, in the federated learning system 1 of this embodiment, each base station 100 performs prescribed learning locally based solely on information obtainable through communication with the terminal 300 connected to it, and each base station 100 generates a local learning model. That is, with Figure 1For example, base station 100a learns based on data obtained through communication with terminals 300a and 300b connected to base station 100a, and base station 100b learns based on data obtained through communication with terminals 300c and 300d connected to base station 100b. Then, each base station 100 sends information related to its generated local learning model to service management device 200. In this way, in federated learning system 1, it is not necessary to send a large amount of information related to terminals 300 from base station 100 to service management device 200, thus preventing communication strain between base station 100 and service management device 200.

[0017] Then, the service management device 200 generates a global learning model based on the information related to the local learning model received from each base station 100, and feeds it back to each base station 100. This prevents the generated local learning model from becoming a model specific to certain information due to insufficient information for learning in base stations 100 with a small number of connected terminals 300.

[0018] It should be noted that terminal 300 is a mobile terminal held by the user. For example, it can be a smartphone, tablet terminal, mobile phone, or other mobile communication module, but it is not limited to these.

[0019] The federated learning system 1 of this embodiment will now be described in detail.

[0020] <Structure> <Structure of Base Station 100> Figure 2 This is a block diagram illustrating a structural example of base station 100. Base station 100 functions as a relay station for communication between ground terminals 300 and is a computer system operating according to a prescribed procedure. Base station 100 can be constructed using a vRAN (virtual radio access network) implemented by a CPU and GPU. Furthermore, in base station 100, the following communication unit 110 can be a RU (radio unit) in 5G communication. Additionally, the functions implemented by control unit 130 can be implemented by a DU (distribution unit) or CU (central unit) in 5G communication, or by a RU. Furthermore, base station 100 can implement an RIC (radio access network intelligent controller) and can operate under the control of the RIC.

[0021] like Figure 2As shown, the base station 100 includes a communication unit 110, a control unit 130, and a storage unit 140. Additionally, the base station 100 may also include an input unit 120 and an output unit 150.

[0022] The communication unit 110 is a communication interface capable of communicating with external devices of the base station 100. The communication unit 110 has the function of communicating with the terminal 300 and the service management unit 200, which are external devices. The communication unit 110 receives signals from the terminal 300 and transmits them to the control unit 130, or sends signals to the terminal 300 according to instructions from the control unit 130. Furthermore, the communication unit 110 transmits information related to the local learning model to the service management unit 200 according to instructions from the control unit 130, or receives information related to the global learning model from the service management unit 200 and transmits it to the control unit 130.

[0023] The input unit 120 has the function of receiving input from operators or the like at the base station 100 and transmitting the input content to the control unit 130. The input unit 120 can be implemented using input devices such as a mouse, keyboard, or touch panel; in the case of voice input, it can be implemented using a microphone. The input unit 120 can, for example, receive specifications for parameters in the local learning model 141 that should be sent to the service management device 200 as information related to the local learning model 141, and transmit them to the control unit 130.

[0024] The control unit 130 is a processor that controls various parts of the base station 100. The control unit 130 can be implemented using a single core or multiple cores. The control unit 130 performs the functions of the base station 100 by executing various programs stored in the storage unit 140 and utilizing various data.

[0025] The control unit 130 relays communication between terminals 300 in the same way as a normal base station 100.

[0026] The control unit 130 has a local learning unit 131, a sending unit 132, and an AI processing unit 133 as functions implemented by the control unit 130.

[0027] The local learning unit 131 learns based on information obtained from communication between the communication unit 110 and each terminal 300. The local learning unit 131 can perform real-time RIC (Regional Information Communication). The learning content of the local learning unit 131 can be preset by the operator managing the federated learning system 1. The local learning unit 131 can perform learning to improve the communication performance between the base station 100 and the terminals 300. For example, the local learning unit 131 can learn the relationship between signals received from the terminals 300 (e.g., SRS (Sounding Reference Signal)) and the current channel information with the terminals 300. That is, the local learning unit 131 generates training data using the signals received from the terminals 300 as explanatory variables and the current channel information with the terminals 300 as target variables, and learns from this data to generate a local learning model 141. The local learning unit 131 stores the generated local learning model 141 in the storage unit 140.

[0028] Furthermore, the local learning unit 131 receives information related to the global learning model from the service management device 200, which acts as a higher-level device, and reflects this information in the local learning model 141. The information related to the global learning model may be information about a portion of the parameters of the local learning model 141. The local learning unit 131 replaces the parameters in the corresponding parts of the local learning model 141 with the parameters transmitted from the service management device 200. Therefore, the local learning model 141 can be modified into a generalized model.

[0029] The local learning unit 131 can communicate with the terminal 300, receive signals from the terminal 300, and perform learning in real time, but is not limited to real time. By learning in real time, the local learning unit 131 can perform reinforcement learning on the local learning model 141, and each time it learns repeatedly, it can grow into a model that is most suitable for the location where the base station 100 is configured.

[0030] The transmitting unit 132 transmits information related to the local learning model 141 generated by the local learning unit 131 to the service management device 200 via the communication unit 110. This information related to the local learning model 141 may be a subset of parameters of the local learning model 141. These parameters may be pre-specified parameters. By transmitting only a subset of the parameters of the local learning model 141, communication strain between the base station 100 and the service management device 200 can be prevented. The information related to the local learning model 141 may be information that establishes a correspondence between information indicating which parameter constitutes the local learning model 141 and its value; the number of parameters can be arbitrary, as long as it includes more than one parameter.

[0031] AI processing unit 133 performs AI processing using a local learning model 141 stored in storage unit 140. The content of AI processing varies depending on the learned content. As described above, if the local learning model 141 is a model that learns the relationship between the signal received from terminal 300 and the channel information between the station and terminal 300, then AI processing unit 133 can perform channel estimation processing based on the transmitted signal received from terminal 300. If the channel with terminal 300 can be estimated, a precoding matrix for terminal 300 can be uniquely determined and beamforming can be performed, thereby improving communication performance with terminal 300.

[0032] The storage unit 140 has the function of storing various programs and data required for the operation of the base station 100. The storage unit 140 can be implemented, for example, using an HDD (Hard Disc Drive), SSD (Solid State Drive), flash memory, etc., but is not limited to these. The storage unit 140 can store various programs and data used to implement the functions to be performed by the base station 100. The storage unit 140 can store programs for learning based on information obtained through communication with the terminal 300. The storage unit 140 can store the local learning model 141 generated by the local learning unit 131.

[0033] The output unit 150 has the function of outputting specified information according to the instructions from the control unit 130. For example, the output unit 150 can output text information or image information; in this case, the output unit 150 is implemented by a monitor installed on or connected to the base station 100. Alternatively, for example, the output unit 150 can output voice information; in this case, the output unit 150 is implemented by a speaker installed on or connected to the information processing device.

[0034] The above is a structural example of base station 100.

[0035] <Structure of Service Management Device 200> Figure 3 This is a block diagram illustrating a structural example of the service management device 200. The service management device 200 is an information processing device (computer system) for implementing a globalized local learning model 141 generated by each base station 100, and may also function as a base station.

[0036] like Figure 3 As shown, the service management device 200 includes a communication unit 210, a control unit 230, and a storage unit 240. Additionally, the service management device 200 may also include an input unit 220 and an output unit 250.

[0037] The communication unit 210 is a communication interface capable of communicating with external devices, such as the service management device 200. The communication unit 210 has the function of communicating with the terminal 300 and the service management device 200, which are external devices. The communication unit 210 receives signals from the terminal 300 and transmits them to the control unit 230, or sends signals to the terminal 300 according to instructions from the control unit 230. Furthermore, the communication unit 210 transmits information related to the local learning model to the service management device 200 according to instructions from the control unit 230, or receives information related to the global learning model from the service management device 200 and transmits it to the control unit 230.

[0038] The input unit 220 has the function of receiving input from operators or other users of the service management device 200 and transmitting the input content to the control unit 230. The input unit 220 can be implemented using input devices such as a mouse, keyboard, or touch panel, and in the case of voice input, it can be implemented using a microphone.

[0039] The control unit 230 is a processor that controls the various parts of the service management device 200. The control unit 230 can be implemented using a single core or multiple cores. The control unit 230 performs the functions of the service management device 200 by executing various programs stored in the storage unit 240 and utilizing various data.

[0040] The control unit 230 relays communication between terminals 300 in the same way as a normal base station 100.

[0041] The control unit 230 has a global learning unit 231 and a feedback unit 232 as functions implemented by the control unit 230.

[0042] The global learning unit 231 generates information 241 related to the global learning model based on information related to the local learning model 141 of the base station 100 transmitted from the communication unit 210. The global learning unit 231 can perform non-real-time RIC (Real-Time Information Conversion). The information 241 related to the global learning model can be obtained by averaging the parameters of the information related to the local learning model 141 transmitted from each base station 100. The global learning unit 231 transmits the generated information 241 related to the global learning model to the feedback unit 232.

[0043] Feedback unit 232 sends (feedback) the information 241 related to the global learning model to each base station 100 via communication unit 210. The information 241 related to the global learning model can be information that establishes a correspondence between information indicating which parameter in the parameters constituting the local learning model 141 corresponds to its value (averaged value), and can be any number of parameters as long as it contains more than one parameter.

[0044] The storage unit 240 has the function of storing various programs and data required for the operation of the service management device 200. The storage unit 240 can be implemented, for example, using an HDD (Hard Disc Drive), SSD (Solid State Drive), flash memory, etc., but is not limited to these. The storage unit 240 can store various programs and data used to implement the functions to be performed by the service management device 200. The storage unit 240 can store programs for generating information 241 related to the global learning model, etc. Furthermore, the storage unit 240 can store information 241 related to the local learning model 141 and information 242 related to the global learning model received from each base station 100.

[0045] The output unit 250 has the function of outputting specified information according to the instructions from the control unit 230. For example, the output unit 250 can output text information or image information; in this case, the output unit 250 is implemented by a monitor installed on or connected to the service management device 200. Alternatively, for example, the output unit 250 can output voice information; in this case, the output unit 250 is implemented by a speaker installed on or connected to the information processing device.

[0046] The above is a structural example of the service management device 200.

[0047] It should be noted that, regarding terminal 300, since it is the same as a typical information processing device and has the same structure as communication terminals such as smartphones and tablets, a detailed explanation of its usage diagram is omitted.

[0048] <Action> From here on, the operation of the base station 100 and service management device 200 in the Federated Learning System 1 will be described.

[0049] Figure 4 This is a flowchart representing an example of the learning action of base station 100.

[0050] like Figure 4 As shown, the communication unit 110 of the base station 100 receives a signal from the terminal 300 (step S401). The communication unit 110 transmits the received signal to the control unit 130.

[0051] The local learning unit 131 performs learning based on the transmitted signal (step S402). The local learning unit 131 generates a local learning model 141 (step S403) and stores it in the storage unit 140.

[0052] The sending unit 132 sends information related to the local learning model 141 to the service management device 200 via the communication unit 110 (step S404). That is, the sending unit 132 extracts the parameters constituting the local learning model 141 and pre-specified parameters, and sends them to the service management device 200. As a result, information 241 related to the global learning model is generated in the service management device 200.

[0053] The communication unit 110 receives information 241 related to the global learning model from the service management device 200 (step S405). The communication unit 110 transmits the received information 241 related to the global learning model to the control unit 130. The information 241 related to the global learning model is information corresponding to the information related to the local learning model 141 sent by this station, and is information averaged by the service management device 200 among multiple base stations 100.

[0054] The local learning unit 131 reflects the parameters represented by the information 241 related to the global learning model into the local learning model 141 (step S406). That is, the local learning unit 131 replaces (overwrites) the parameters of the corresponding parts of the local learning model 141 with the transmitted parameters. As a result, even if the number of terminals 300 communicating with the base station 100 is small, a generalized local learning model 141 can be obtained, which can improve the accuracy of AI processing.

[0055] The AI ​​processing unit 133 takes the information about the terminal 300 transmitted from the communication unit 110 as input and performs AI processing using the updated local learning model 141 (step S407). As described above, if the local learning model 141 is a model that has learned the relationship between the signal received from the terminal 300 and the channel information for the terminal 300 at that time, it can perform AI processing by estimating the channel information between the terminal 300 and the terminal. The estimated channel information can be used in subsequent communications with the terminal 300; for example, it can be used to generate a precoding matrix for beamforming of the terminal 300.

[0056] The above is an example of the operation of base station 100 in the implementation method.

[0057] Figure 5 This is a flowchart representing an example of the learning action of the service management device 200.

[0058] like Figure 5As shown, the communication unit 210 of the service management device 200 receives information related to the local learning model 141 generated in each base station 100 (step S501). The information related to the local learning model 141 may be information about parameters that constitute part of the local learning model 141. The communication unit 210 transmits the received information related to the local learning model 141 to the control unit 230.

[0059] The global learning unit 231 of the control unit 230 averages the parameters corresponding to the information related to the local learning model 141 transmitted from multiple base stations 100, thereby generating information 241 related to the global learning model (step S502).

[0060] Feedback unit 232 feeds back information 241 related to the global learning model generated by global learning unit 231 to each base station 100 via communication unit 210 (step S503), and ends processing.

[0061] The above is an example of the operation of the service management device 200 in the implementation method.

[0062] Summary As described above, base station 100 learns in real time using signals received from its respective connected terminals to generate local learning models. Therefore, it is not necessary to send a large amount of information for learning from base station 100 to service management device 200, thereby reducing the possibility of communication strain between base station 100 and service management device 200. In addition, since each base station 100 only uses information it holds to learn, it may not be able to obtain enough data, resulting in long learning times, or it may generate local learning models 141 without sufficient learning, leading to a decrease in estimation accuracy during operation. However, by feeding back information 241 related to the global learning model from service management device 200 and reflecting it in each local learning model 141, the accuracy of the estimation results of local learning models 141 can be prevented from decreasing. In addition, in each base station 100, generalization is achieved through information 241 related to the global learning model. At the same time, through continuous learning in real time, a local learning model 141 suitable for the location of the base station 100 can be generated. Furthermore, AI processing can be performed using the local learning model 141 suitable for the regional environment, thereby increasing the possibility of performing processing most suitable for the environment in which the base station 100 is located.

[0063] <Variation Example> The service management device 200 and base station 100 described in the above embodiments are not limited to the above embodiments, and can of course be implemented by other methods. Hereinafter, various modifications will be described.

[0064] (1) In the above embodiment, the base station 100 is configured to send information of pre-specified parameters as information related to the local learning model 141, and the service management device 200 is configured to send the average value of the parameters as information related to the global learning model to each base station 100, and reflect it in the local learning model 141. However, the base station 100 may also be configured to send the generated local learning model 141 itself to the service management device 200. Then, the service management device 200 may also average the entire local learning model 141 to generate information 242 related to the global learning model. In this case, the information 242 related to the global learning model becomes the global learning model itself. Then, the feedback unit 232 may feed back the generated global learning model itself as information 242 related to the global learning model to each base station 100.

[0065] (2) In the above embodiment, the local learning unit 131 can be implemented by an external information processing device that can communicate with the base station 100 and receive information obtained through communication with the terminal 300.

[0066] (3) In the above embodiment, an example was shown where the global learning unit 231 of the service management device 200 performs learning at certain times without real-time. However, the timing of the learning by the global learning unit 231 can be appropriately changed. For example, when the federated learning system 1 starts learning, the global learning unit 231 can perform learning at a high frequency, considering that the information used for learning cannot be fully accumulated. On the other hand, after the system starts running and a certain period of time has passed, the possibility of a certain degree of stable communication being able to be performed in each base station 100 is high. In this case, even if the information related to the global learning model is not reflected to each base station 100, it is not expected that there will be a large change. Therefore, the global learning unit 231 can reduce the frequency of learning. That is, the global learning unit 231 can perform processing that relatively reduces the frequency of learning (the frequency of averaging specific parameters of the information group related to the local learning model 141) over time. As a result, the processing load of the global learning unit 231 of the service management device 200 and the local learning unit 131 of the base station 100 can be reduced over time.

[0067] (4) In the above embodiments, as a reference Figure 4 The actions of the base station 100 described herein, the processing of steps S401 to S406 and the processing of step S407 can be performed in parallel, and the processing of step S407 does not necessarily have to be performed after the processing of step S406.

[0068] (5) In the above embodiment, it was explained that the local learning model 141 is a model that learns the relationship between information obtained through communication with the terminal 300 and the channel between the local learning model 141 and the terminal 300 at that time. As mentioned above, the local learning model 141 learns different content depending on the content of the AI ​​processing to be performed, and the processing is also different. As another example, the local learning model 141 may be a model that learns the correspondence between signals from the terminal 300 and communication settings (communication frequency, communication timing, etc.) applicable to the terminal 300 at that time. In this case, the AI ​​processing can take the signals from the terminal 300 as input and estimate the communication settings suitable for the terminal 300.

[0069] In addition, the local learning model 141 does not need to learn training data with the signal of one terminal 300 as the explanatory variable, but can also learn training data with the signals of multiple terminals 300 as the explanatory variables.

[0070] Thus, the local learning model 141 can be any model as long as it is configured to perform the estimation processing required for the intended purpose. Furthermore, at this time, it is sufficient as long as the base station 100 itself can acquire the data that serves as the explanatory variable and the data that serves as the target variable, which can be used as training data.

[0071] (6) The program disclosed herein, which allows the base station 100 to create a local learning model 141 based on information from the terminal 300, and the program that allows the service management device 200 to generate information related to the global learning model, can be provided in a state of being stored in a computer-readable storage medium. The storage medium can store the program in a "non-transitory tangible medium". The storage medium can include any suitable storage medium such as HDD, SSD, or two or more suitable combinations thereof. The storage medium can be volatile, non-volatile, or a combination of volatile and non-volatile. It should be noted that the storage medium is not limited to these examples; it can be any device or medium as long as it can store the program.

[0072] It should be noted that the base station 100 and service management device 200 can, for example, read and execute a program stored in a storage medium to implement the functions of the multiple functional units shown in each embodiment. Furthermore, this program can be provided to the base station 100 and service management device 200 via any transmission medium (communication network, broadcast wave, etc.). The base station 100 and service management device 200 can, for example, execute a program downloaded via the Internet to implement the functions of the multiple functional units shown in each embodiment. This program can be executed by the base station 100, service management device 200, etc.

[0073] It should be noted that the program can be implemented using scripting languages ​​such as ActionScript and JavaScript (registered trademark), object-oriented programming languages ​​such as Objective-C and Java (registered trademark), and markup languages ​​such as HTML5, but is not limited to these.

[0074] At least a portion of the processing in the base station 100 and service management device 200 can be implemented using cloud computing, which consists of one or more computers. Furthermore, each functional unit of the base station 100 can be implemented by one or more circuits that implement the functions described in the above embodiments, or multiple functional units can be implemented by a single circuit.

[0075] (7) The various methods and processes shown in the above embodiments and variations can be appropriately combined within the scope of achieving the purpose of performing AI processing by federated learning between the base station 100 and the service management device 200.

[0076] (8) According to the various methods of this disclosure described above, by being able to perform AI processing, the AI ​​processing can realize fast and high-precision communication between base stations and terminals, and further between terminals, which can help achieve goal 9 of the Sustainable Development Goals (SDGs) "Building industry, innovation and infrastructure". Explanation of reference numerals in the attached figures

[0077] 100, 100a, 100b: Base station; 110: Communication unit; 120: Input unit; 130: Control unit; 131: Local learning unit; 132: Transmission unit; 133: AI processing unit; 140: Storage unit; 150: Output unit; 200: Service management device; 210: Communication unit; 220: Input unit; 230: Control unit; 231: Global learning unit; 232: Feedback unit; 240: Storage unit; 250: Output unit; 300, 300a, 300b, 300c, 300d: Terminal.

Claims

1. A federated learning system, the federated learning system comprising multiple base stations and a host device capable of communicating with each of the multiple base stations, characterized in that, The base station has the following features: The learning unit learns communication information related to communication between the terminal and the learning unit, and generates a local learning model related to the communication through the real-time wireless access network intelligent controller, i.e., the real-time RIC; and The sending unit transmits information related to the local learning model to the host device. The host device includes: The receiving unit receives information related to the local learning model; The generation unit generates an ensemble learning model based on the local learning models of the multiple base stations through non-real-time RIC. as well as The feedback unit feeds back information related to the ensemble learning model generated by the generation unit to the base station.

2. The federated learning system according to claim 1, characterized in that, The information related to the local learning model is a pre-specified portion of the parameter information of the local learning model. The ensemble learning model is generated by averaging the parameter information sent from the multiple base stations respectively. The feedback unit feeds back the averaged parameter information to the base station. The learning unit replaces the parameters in the local learning model that correspond to the parameter information with the averaged parameters.

3. The federated learning system according to claim 2, characterized in that, As time goes by, the generation unit reduces the frequency of generating ensemble learning models through the non-real-time RIC.

4. A federated learning method, which is a federated learning method in a federated learning system comprising multiple base stations and a host device capable of communicating with each of the multiple base stations, characterized in that, Perform the following steps: The learning process involves the base station learning communication information related to communication between itself and the terminal, and generating a local learning model related to communication through a real-time wireless access network intelligent controller, i.e., a real-time RIC. In the transmission step, the base station sends information related to the local learning model to the host device; In the receiving step, the host device receives information related to the local learning model; In the generation step, the host device generates an ensemble learning model based on the local learning models of the multiple base stations through non-real-time RIC. as well as The feedback step feeds back information related to the ensemble learning model generated in the generation step to the base station.

Images