Channel environment prediction method and apparatus, management device, storage medium, and computer program product

Abstract

The application discloses a channel environment prediction method and device, a management device, a storage medium and a computer program product. The method comprises the following steps: a management device determines first information, the management device is used for managing a plurality of access network devices at least, the first information represents that a first model locally of one or more first access network devices cannot be trained and / or a channel environment predicted by using the first model locally cannot optimize network performance, and the first model is used for predicting the channel environment; a corresponding second model is provided for each of the one or more first access network devices, the second model is used for predicting the channel environment, and the second model is obtained by federated learning of the plurality of access network devices managed by the management device.

Description

Technical Field

[0001] This application relates to the field of wireless technology, and in particular to a channel environment prediction method, apparatus, management device, storage medium, and computer program product. Background Technology

[0002] In related technologies, the channel environment at future moments can be predicted based on artificial intelligence (AI) models, yielding the predicted channel environment. Base stations can then adjust resource allocation based on the predicted channel environment, such as adjusting the modulation scheme and coding rate of wireless link transmission, effectively improving network performance at future moments (e.g., after the base station has adjusted resource allocation).

[0003] However, in practical applications, there is currently no effective solution for obtaining a model that can accurately predict the channel environment. Summary of the Invention

[0004] To address the related technical issues, embodiments of this application provide a channel environment prediction method, apparatus, management device, storage medium, and computer program product.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] This application provides a channel environment prediction method applied to a management device, wherein the management device is used to manage at least multiple access network devices, and the method includes:

[0007] Determine first information, the first information being characterized by the fact that a first model local to one or more first access network devices cannot be trained and / or that the channel environment predicted by the local first model cannot optimize network performance, the first model being used to predict the channel environment;

[0008] A corresponding second model is provided for each of one or more first access network devices. The second model is used to predict the channel environment and is obtained by federated learning using multiple access network devices managed by the management device.

[0009] The method in the above scheme further includes:

[0010] Obtain second information about multiple access network devices, the second information including relevant information about the access network devices;

[0011] Using the acquired second information, the multiple access network devices are divided into M groups, each group containing multiple access network devices, where M is an integer greater than or equal to 1;

[0012] For each group, an aggregation device and multiple training devices are determined. The training devices are used to train the second model using local data and obtain training results. The training devices belong to multiple access network devices corresponding to the group. The aggregation device is used to aggregate the training results obtained by the multiple training devices to obtain the trained second model and provide the trained second model to each first access network device included in the group. Federated learning is performed using the aggregation device and multiple training devices.

[0013] In the above scheme, the step of dividing the multiple access network devices into M groups using the acquired second information includes:

[0014] For each of the plurality of second pieces of information, feature extraction is performed on the second piece of information to obtain the feature information of the access network device;

[0015] Based on the similarity of feature information of multiple access network devices, the multiple access network devices are divided into M groups.

[0016] In the above scheme, the second information includes relevant information about the local data of the access network device, and the step of determining multiple training devices for each group includes:

[0017] For each group, multiple training devices are determined by utilizing the computing power resource information and / or relevant information of the local data of the multiple access network devices corresponding to the group.

[0018] In the above scheme, determining an aggregation device for each group includes:

[0019] For each group, the aggregation device is determined by utilizing the computing power resource information and / or transmission latency information related to multiple access network devices corresponding to the group, the management device, and one or more other management devices besides the management device.

[0020] In the above scheme, obtaining the second information of multiple access network devices includes:

[0021] Send a fourth message to the plurality of access network devices, the fourth message being used to request the reporting of the second message; receive the second message reported by the plurality of access network devices;

[0022] or,

[0023] Subscribe to the second information of the multiple access network devices.

[0024] The method in the above scheme further includes:

[0025] The fifth piece of information is determined, which indicates that one or more second access network devices are unable to optimize network performance using the channel environment predicted by the second model;

[0026] The second model is retrained to obtain a retrained second model; a retrained second model is provided for each of the one or more second access network devices.

[0027] The method in the above scheme further includes:

[0028] Obtain the prediction performance of the plurality of access network devices in predicting the channel environment using the second model; and designate the access network devices with prediction performance lower than a first threshold as the second access network devices.

[0029] or,

[0030] Receive sixth information sent by one or more access network devices, the sixth information indicating that the prediction performance of the channel environment using the second model is lower than a second threshold; designate the access network device corresponding to the sixth information as the second access network device.

[0031] In the above scheme, retraining the second model includes:

[0032] The second model is trained by reusing the aggregation device and multiple training devices;

[0033] or,

[0034] Using the acquired second information, the multiple access network devices are re-divided into N groups, where N is an integer greater than or equal to 1; for each re-divided group, an aggregation device and multiple training devices are determined; the second model is trained using the aggregation device and multiple training devices.

[0035] In the above scheme, N is less than M when the second model is overfitted; or N is greater than M when the second model is underfitted.

[0036] This application embodiment also provides a channel environment prediction device, disposed in a management device, the management device being used to manage at least a plurality of access network devices, including:

[0037] A determining unit is configured to determine first information, the first information representing that a first model local to one or more first access network devices cannot be trained and / or the channel environment predicted by the local first model cannot optimize network performance, the first model being used to predict the channel environment;

[0038] A providing unit is configured to provide a corresponding second model for each of one or more first access network devices, the second model being used to predict the channel environment, and the second model being obtained by federated learning using multiple access network devices managed by the management device.

[0039] This application also provides a management device, including: a processor and a memory for storing computer programs capable of running on the processor.

[0040] When the processor runs the computer program, it executes the steps of any of the above methods.

[0041] This application also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of any of the above methods.

[0042] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above methods.

[0043] The channel environment prediction method, apparatus, management device, storage medium, and computer program product provided in this application embodiment include: a management device determining first information, the management device being used to manage at least multiple access network devices; the first information indicating that a first model local to one or more first access network devices cannot be trained and / or the channel environment predicted using the local first model cannot optimize network performance; the first model being used to predict the channel environment; and providing a corresponding second model for each of the one or more first access network devices, the second model being used to predict the channel environment, the second model being obtained by federated learning using the multiple access network devices managed by the management device. The solution provided in this application embodiment involves an access network device first using a local model to predict the channel environment. When a management device used to manage multiple access network devices determines that one or more access network devices have difficulty training a local model and / or the performance of the local model is insufficient to meet the requirements (e.g., the predicted channel environment cannot optimize network performance), the management device provides these access network devices with a model obtained through federated learning from multiple access network devices. Thus, when the access network devices use the model provided by the management device to predict the channel environment, because the model provided by the management device is trained using multiple access network devices through federated learning, that is, it has been fully trained using the data and computing resources of multiple access network devices, the model provided by the management device can achieve better channel environment prediction performance (e.g., accuracy), and the predicted channel environment can better meet the requirements for optimizing network performance. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the channel environment prediction method according to an embodiment of this application;

[0045] Figure 2 This is a schematic diagram illustrating the application example of predicting future channel matrix data based on historical channel matrix data.

[0046] Figure 3 This is a flowchart illustrating a channel environment prediction method based on federated learning, which serves as an application example of this application.

[0047] Figure 4 This is a schematic diagram of the channel environment prediction device according to an embodiment of this application;

[0048] Figure 5 This is a schematic diagram of the management device structure according to an embodiment of this application. Detailed Implementation

[0049] The present application will now be described in further detail with reference to the accompanying drawings and embodiments.

[0050] In wireless communication systems such as 4G Long Term Evolution (LTE) and 5G New Radio (NR), network devices can typically allocate resources (i.e., wireless resources) based on wireless channel measurement information between the network side and the terminal.

[0051] For example, in a wireless communication system, access network equipment (such as a base station) can utilize Adaptive Modulation and Coding (AMC) technology for resource allocation. This means that based on changes in wireless channel measurement information (also understood as wireless channel measurement values) between the network side and the terminal, the modulation scheme and coding rate of the wireless link transmission are adaptively adjusted. Specifically, when the wireless channel measurement information indicates a poor current channel environment, a lower-order modulation and coding scheme (MCS) can be selected to improve the signal's anti-interference capability and reduce the error rate. When the wireless channel measurement values ​​indicate a good current channel environment, a higher-order MCS can be selected to maximize the wireless link transmission efficiency between the network side and the terminal while ensuring the reliability of the wireless link transmission.

[0052] However, when adjusting (or determining) the MCS using traditional AMC technology, the wireless channel measurement information acquired by the access network equipment is often outdated. This means that the channel environment (channel quality, channel conditions, etc.) represented by the acquired wireless channel measurement information may differ from the channel environment at the time of wireless transmission after MCS adjustment. Therefore, it is difficult to accurately reflect the channel environment at the moment of wireless transmission using the adjusted MCS (e.g., the time of uplink data transmission). Consequently, in scenarios with significant channel environment fluctuations, such as when terminals move and / or when the network deploys massive MIMO for communication, an MCS adjusted using outdated wireless channel measurement information may fail to meet actual service requirements, leading to problems such as high network bit error rate and decreased throughput performance.

[0053] In related technologies, on the one hand, a scheme is proposed to predict the channel environment using AI algorithm-related models (hereinafter referred to as AI models). Specifically, the channel environment is first predicted using AI models to obtain the predicted future channel environment, and then resource optimization configuration (such as adjusting MCS) is performed based on the predicted channel environment so that the optimized resource configuration can effectively optimize network performance.

[0054] However, in practical applications, schemes that use AI models to predict the channel environment usually require a lot of computing resources and large-scale datasets. When a single access network device trains an AI model (such as for machine learning training), it may face problems such as insufficient computing power and / or insufficient data, which makes it difficult to train the AI ​​model and / or the performance of the trained AI model is poor (such as low accuracy in predicting the channel environment).

[0055] On the other hand, federated learning is a learning method that employs a distributed architecture. This architecture includes an aggregation device (which can also be understood as an aggregation node, a global model aggregation node, or a server) and multiple training devices (which can also be understood as training nodes or clients). Each training device can utilize its own computing and data resources to train the model, obtaining its own training results. The aggregation device then integrates the training results from each device to obtain the global model. Compared to schemes that rely on a single device for model training, federated learning requires less computing power and data resources from each training device. Therefore, even when training device resources are limited, federated learning can still yield a high-performance model.

[0056] In related technologies, training a model using federated learning typically involves the following steps:

[0057] Step a: The aggregation device acquires the initial model (which can also be understood as the model to be trained);

[0058] Step b: The aggregation device sends the initial model to multiple training devices respectively;

[0059] Step c: After receiving the initial model, each training device trains the model using local data, and after training is completed, sends the trained model and / or the updated information of the trained model compared with the initial model (such as gradients, model parameters, etc.) to the aggregation device.

[0060] Step d: The aggregation device receives the trained model and / or the updated information of the trained model compared with the initial model reported by each training device (hereinafter referred to as the reported information). Then, it uses all the received reported information and combines the aggregation algorithm (which can also be understood as the fusion algorithm, such as averaging or weighted averaging) to update (or optimize or fine-tune) the initial model to obtain the updated initial model.

[0061] In practical applications, steps b through d can be repeated to perform multiple rounds of training and improve model performance. The aggregation device can stop repeating the process and proceed to step e once the number of repetitions reaches a preset threshold and / or the performance of the updated initial model meets preset requirements.

[0062] Step e: The aggregation device uses the updated initial model as the trained model and deploys the trained model to relevant devices so that the relevant devices can use the trained model to perform relevant business.

[0063] As can be seen from the above description, federated learning allows multiple training devices to collaboratively train a single model. Each training device uses local data for model training, while the aggregation device combines the training results from multiple devices to obtain the trained model. Specifically, this offers the following advantages:

[0064] 1) Multiple training devices can train models without sharing training data (i.e., multiple training devices do not need to share training data), which reduces the communication transmission cost and time cost of transmitting training data (which can also be understood as the raw training data);

[0065] 2) Simultaneous model training using multiple training devices saves training time;

[0066] 3) The computing and data resources of multiple training devices can be used simultaneously for model training. In this way, even if the resources of each training device are limited, the model can be trained sufficiently to obtain a model with better performance.

[0067] 4) In the process of training the model through federated learning, the data characteristics of different training devices are considered, and the data diversity and comprehensiveness are high. Therefore, the generalization ability and comprehensiveness of the model trained through federated learning are also better.

[0068] Based on this, in various embodiments of this application, the access network device first uses a local model to predict the channel environment. When the management device for managing multiple access network devices determines that one or more access network devices have difficulty training a local model and / or the performance of the local model is insufficient to meet the requirements (e.g., the predicted channel environment cannot optimize network performance), it provides these access network devices with a model obtained by federated learning from multiple access network devices. When the access network devices use the model provided by the management device to predict the channel environment, because the model provided by the management device is trained by multiple access network devices through federated learning, that is, it has been fully trained using the data and computing resources of multiple access network devices, the model provided by the management device can achieve better channel environment prediction performance (e.g., accuracy), and thus the predicted channel environment can better meet the requirements for optimizing network performance.

[0069] This application provides a channel environment prediction method applied to a management device, wherein the management device is used to manage at least multiple access network devices, such as... Figure 1 As shown, the method includes:

[0070] Step 101: Determine first information, the first information representing that the first model of one or more first access network devices cannot be trained and / or the channel environment predicted by the local first model cannot optimize network performance, the first model being used to predict the channel environment;

[0071] Step 102: Provide a corresponding second model for each of the one or more first access network devices. The second model is used to predict the channel environment and is obtained by federated learning using the multiple access network devices managed by the management device.

[0072] In practical applications, the management device may include access network devices or upper-layer network devices. For example, the access network device may include a base station (e.g., a gNB), and the upper-layer network device may include Operations, Administration and Maintenance (OAM) equipment, a Near Real Time Radio Access Network Intelligent Controller (NRT RIC), or a Non-Real-Time Radio Access Network Intelligent Controller (Non-RT RIC), etc. The management device may also be referred to as an intelligent network device; however, this application does not limit the name or specific implementation of the management device.

[0073] In practical applications, the terminal can measure the channel environment, obtain measurement results, determine the channel environment based on the measurement results, and report relevant information about the channel environment (also referred to as channel environment information) to the access network device. After receiving the channel environment information, the access network device can select an MCS suitable for the channel environment (also understood as matching the channel environment) and conduct subsequent communication with the terminal according to the selected MCS (also understood as conducting future communication) to optimize communication network resources (such as maximizing transmission efficiency). The channel environment information may include a Channel Quality Indicator (CQI); the terminal may be referred to as a User Equipment (UE), a user, etc., which is not limited in this embodiment. Here, the above process can also be understood as resource scheduling optimization based on wireless environment prediction.

[0074] However, there is a time delay between the terminal's channel environment measurement and the subsequent communication between the access network device and the terminal based on the selected MCS. This may result in a mismatch between the MCS selected based on the terminal's measurement results and the channel environment when the access network device and the terminal communicate subsequently. This can also be understood as the measurement results reported by the terminal being outdated.

[0075] For example, suppose that the terminal measures the non-zero-power channel state reference signal (NZP-CSI-RS) and channel state information interference measurement (CSI-IM) related to the access network equipment at the time domain position of the nth transmission time interval (TTI) (which can also be understood as tti n), and obtains the measurement results; the terminal uses the measurement results to determine (or estimate) the signal strength S(n) (which can also be understood as useful signal strength) and noise strength I(n) (which can also be understood as interference strength) of the subcarrier level, or resource block (RB) level, or RB group (RBG) level (hereinafter referred to as subcarrier / RB / RBG level); the terminal can use the signal strength S(n) and noise strength I(n) of the subcarrier / RB / RBG level to determine the signal-to-noise ratio SINR(n) of the subcarrier / RB / RBG level, and the specific formula can be expressed as formula (1):

[0076]

[0077] in, This represents the intensity of Gaussian white noise. The process by which the terminal determines the signal-to-noise ratio (SINR) (n) of the subcarrier / RB / RBG level using the measurement results can also be called channel estimation.

[0078] After determining the signal-to-noise ratio (SINR)(n) at the subcarrier / RB / RBG level, the terminal can use the SINR(n) to determine the CQI corresponding to the SINR(n) (for example, by looking up a table or other methods), and report the determined CQI to the access network device. After receiving the CQI reported by the terminal at the (n+Δ)th TTI (or tti n+Δ), the access network device uses the CQI to determine the channel environment between the access network device and the terminal, as well as the MCS suitable for that channel environment. When the access network device and the terminal perform the next service data transmission (which may include downlink data service transmission or uplink service transmission) at the (n+Δ+μ)th TTI (or tti n+Δ+μ), data transmission can be performed based on the determined MCS. However, there is a time delay of Δ+μ TTI between the time domain location of the service data transmission and the time domain location of the measurement reference signal. This may cause the MCS determined by the access network equipment to be mismatched with the channel environment when data transmission actually takes place. For example, the terminal may move within the time period corresponding to the delay, causing changes in the channel environment. Therefore, when using a determined MCS for data transmission, it may be difficult to effectively improve network performance.

[0079] To address the aforementioned issues, an AI-based channel environment prediction model (also known as an AI model) can be deployed on the terminal side. After measuring the reference signal, the terminal can use the deployed model to predict the future channel environment (specifically, the channel environment during service data transmission between the access network device and the terminal after selecting an MCS), obtaining predicted channel environment information (such as predicted CQI) and reporting this information to the access network device. Upon receiving the predicted channel environment information, the access network device can use it to determine the corresponding MCS and then use that MCS to transmit service data with the terminal. Therefore, since the predicted channel environment matches the actual channel environment during service data transmission, the access network device can determine the MCS that matches the actual channel environment based on the predicted information. Using the determined MCS for service data transmission effectively improves network performance.

[0080] In practical applications, compared to the method of determining the channel environment based on the measurement results of reference signals, the method of determining the channel environment using AI models requires a higher computational cost (i.e., consumes more computing resources). In other words, if the channel environment between all access network devices and terminals is obtained through AI model prediction, it will require a large amount of computing resources. Therefore, the management device can manage the method of determining the channel environment between each access network device and terminal, reducing resource consumption while ensuring the accuracy of the channel environment determination. The management device's management of the method of determining the channel environment between each access network device and terminal may specifically include the following steps:

[0081] Step 1: The management device sends subscription request information to multiple access network devices under its management (which can also be understood as access network devices within the service range or all access network devices connected to the management device). The subscription request information is used to request subscription to network performance indicators.

[0082] The network performance metrics are used to describe the network performance of the access network devices and / or terminals (i.e., the terminals served by the access network devices). Specifically, the network performance metrics may include one or more of the following (one or more can also be understood as at least one): Instantaneous Block Error Rate (IBLER), throughput, latency, etc. The subscription request information may include reporting trigger conditions, such as requiring the access network device to report subscription information to the management device when the network performance metrics are below a preset threshold.

[0083] Step 2: After receiving the subscription request information, the access network device performs corresponding configuration and detects the network performance indicators related to the subscription request information; when the detected network performance indicators do not meet the reporting trigger conditions (i.e., the network performance is relatively good), the terminal determines the channel environment based on the measurement results of the reference signal; when the detected network performance indicators meet the reporting trigger conditions (i.e., the network performance is relatively poor), step 3 is executed.

[0084] Step 3: The access network device reports subscription information to the management device; wherein, the subscription information may include the identification information of the access network device (such as the access network device ID) and / or the network performance indicators of the channel between the access network device and the terminal;

[0085] Step 4: The management device sends an instruction message to the access network device, the instruction message being used to instruct the use of AI model prediction to determine the channel environment between the access network device and the terminal.

[0086] As can be seen from the above description, when the network performance between the access network device and the terminal is good (i.e., the network performance indicators are good), the terminal determines the channel environment based on the measurement results of the reference signal, without using AI for channel environment prediction, thereby saving computing resources. At the same time, when the network performance between the access network device and the terminal is poor (i.e., the network performance indicators are poor), the management device can instruct the use of AI model prediction to determine the channel environment between the access network device and the terminal, thereby ensuring the accuracy of the determined channel environment and enabling the access network device to select an MCS that is more suitable for the actual channel environment, so as to effectively improve network performance.

[0087] In practical applications, when using AI model prediction to determine the channel environment between the access network device and the terminal, the access network device can acquire a trained AI model (i.e., the trained AI model) and distribute it to the terminal so that the terminal can deploy the AI ​​model and use it to predict the channel environment. Specifically, the access network device can acquire the trained AI model in one of the following ways:

[0088] The first method: The access network device (specifically including the access network device that communicates with the terminal) uses local data to train the AI ​​model and obtain the trained AI model;

[0089] The second approach involves federated learning using an aggregation device and multiple training devices (including access network devices) to obtain a trained AI model; the access network device receives the trained AI model from the aggregation device.

[0090] Among them, when the first method is used to obtain the trained AI model, the AI ​​model is trained on a single access network device, which has a lower computational cost but a higher resource requirement for the single access network device; when the second method is used to obtain the trained AI model, the AI ​​model is trained on multiple access network devices, which has a higher computational cost but a lower resource requirement for each access network device.

[0091] Based on this, when the management device instructs the use of AI model prediction to determine the channel environment between the access network device and the terminal, the access network device can preferentially use the first method to obtain the trained AI model to save computing resources; and when the access network device has difficulty obtaining the AI ​​model through the first method (e.g., unable to perform model training), and / or when the AI ​​model obtained by the first access network device through the first method is difficult to meet performance requirements, the access network device uses the second method to obtain the trained AI model to ensure accurate prediction of the channel environment and effective optimization of network performance.

[0092] Specifically, after the access network device receives the instruction information sent by the management device, it can train an AI model using local data according to the first method described above, thereby obtaining a trained AI model. The process of the access network device training the AI ​​model using local data can include: the access network device receiving channel environment-related information (such as channel matrix information, interference information, etc.) reported by the terminal, and using the received channel environment-related information and historical data recorded by the access network device as local data; the access network device using the local data to train a first model for predicting the channel environment; and after training, the access network device compressing or pruning the trained first model based on the terminal's inference capabilities (such as computing power, storage capacity, etc.) (specifically, this may include removing redundant or unimportant parameters from the model), obtaining a terminal model, and distributing the terminal model to the terminal; after receiving and deploying the terminal model, the terminal can use the terminal model to predict the channel environment. Here, the first model may specifically include a model using a Long Short-Term Memory (LSTM) architecture. The embodiments of this application do not limit the name and specific implementation of the first model. The process by which the access network device trains the first model, uses the first model to determine the terminal model, and sends the terminal model to the terminal, and the terminal uses the terminal model to predict the channel environment, can also be called an intelligent channel environment prediction scheme based on AI algorithms.

[0093] In practical applications, when the access network device trains the first model using local data, one or more of the following problems may occur, resulting in the access network device's local first model being unable to be trained and / or the channel environment predicted by the local first model being unable to optimize network performance:

[0094] 1) The access network equipment has insufficient computing power resources (which can also be understood as insufficient computing power conditions), making it difficult to support the training of the first model;

[0095] 2) The access network equipment has limited power resources (which can also be understood as limited power consumption), making it difficult to support the training of the first model;

[0096] 3) The performance of the first model trained by the access network device is poor, for example, the prediction accuracy of the channel environment is lower than a preset threshold;

[0097] 4) When the access network device trains the first model, the convergence time is too long, resulting in a low prediction accuracy of the first model.

[0098] As can be seen from the above description, when the access network device has one or more of the aforementioned problems, the access network device may have difficulty obtaining the first model through the first method, and / or the first model obtained by the access network device through the first method may be insufficient to meet performance requirements. In this case, the access network device may obtain the trained AI model through the second method.

[0099] Of course, when the access network device determines that the performance of the model obtained through the second method is significantly better than that of the first model (i.e., there is a significant performance gain), the access network device may also choose to obtain the trained AI model through the second method.

[0100] In practical applications, the management device can use multiple access network devices under its management to perform federated learning to obtain a second model (i.e., a trained AI model) for predicting the channel environment, and provide the second model to the access network device when the access network device determines that it has obtained the second model through the second method described above.

[0101] Specifically, the access network device can trigger the management device to perform a federated learning process using multiple managed access network devices (hereinafter referred to as the federated learning process), or the management device can trigger the federated learning process. The access network device can trigger the federated learning process by executing the following steps:

[0102] Step 1: The access network device sends a notification message to the management device to inform the management device that the access network device has one or more of the following conditions (one or more can also be understood as at least one):

[0103] The local first model cannot be trained;

[0104] The channel environment predicted using the local first model cannot optimize network performance;

[0105] Step 2: After receiving the notification information, the management device determines to provide the second model for the access network device (i.e., the access network device corresponding to the notification information);

[0106] Here, step 2 can also be understood as step 101 above, that is, determining the first information; the access network device that provides the second model can also be called the first access network device, that is, the first model of the first access network device cannot be trained and / or the channel environment predicted by the local first model cannot optimize network performance; at the same time, the first information can also be called trigger information, the first information is used to trigger the federated learning process, and the name of the first information is not limited in this application embodiment.

[0107] Step 3: After determining that the second model is provided for the access network device, the management device uses the multiple access network devices under its management to perform federated learning (i.e., perform a federated learning process) to obtain the trained second model;

[0108] Here, step 2 can also be understood as step 102 above.

[0109] Step 4: The aggregation device related to federated learning provides the second trained model to the access network device.

[0110] Meanwhile, the management device can trigger the federated learning process by performing the following steps:

[0111] Step 1: The management device acquires relevant information (such as computing resources, network resources, model performance, etc.) of the multiple access network devices it manages;

[0112] In practical applications, the method by which the management device obtains relevant information about the access network device can be set according to actual needs, such as subscribing to relevant information about the access network device. This application embodiment does not limit this.

[0113] Step 2: The management device uses all the relevant information it has acquired to determine which access network devices cannot have their local first model trained and / or whose channel environment predicted by their local first model cannot optimize network performance; that is, to determine which access network devices belong to the first access network devices; if the number of the determined first access network devices reaches a preset threshold, the first information is determined and step 3 is executed.

[0114] Step 3: The management device uses the multiple access network devices it manages to perform federated learning (i.e., perform the federated learning process) to obtain the trained second model;

[0115] Step 4: The aggregation device related to federated learning provides the second trained model to the access network device.

[0116] As can be seen from the above description, after the management device determines the first information, i.e., after triggering the federated learning process, it can perform federated learning using multiple managed access network devices to obtain a trained second model. In this process, the management device can first determine which access network devices to use for federated learning, i.e., determine the aggregation device and the training device, and then use the determined aggregation device and training device to perform federated learning and obtain the trained second model.

[0117] The following describes the process for determining the aggregation and training devices related to federated learning in the management equipment.

[0118] In practical applications, the multiple access devices managed by the management device may have significant differences, such as differences in channel conditions, computing resources, and communication resources (which can also be understood as connection resources).

[0119] Based on this, the management device can acquire channel environment prediction-related information (such as channel matrix information, computing resources, communication resources, terminal computing resources, interference information, etc.) from multiple access network devices. Then, the management device can group these multiple access network devices according to the similarity of their channel environment prediction-related information, resulting in one or more groups (one or more can also be understood as at least one). For each group, a federated learning aggregation device and a training device are determined. Subsequently, the management device can perform federated learning using the aggregation device and training device corresponding to each group to obtain a second model for that group. Because the channel environment prediction-related information of the multiple access network devices in each group is similar, the access network devices in that group can achieve better prediction accuracy when using the second model corresponding to that group for channel prediction.

[0120] Specifically, in one embodiment, the method may further include:

[0121] Obtain second information about multiple access network devices, the second information including relevant information about the access network devices;

[0122] Using the acquired second information, the multiple access network devices are divided into M groups, each group containing multiple access network devices, where M is an integer greater than or equal to 1;

[0123] For each group, an aggregation device and multiple training devices are determined. The training devices are used to train the second model using local data and obtain training results. The training devices belong to multiple access network devices corresponding to the group. The aggregation device is used to aggregate the training results obtained by the multiple training devices to obtain the trained second model and provide the trained second model to each first access network device included in the group. Federated learning is performed using the aggregation device and multiple training devices.

[0124] In practical applications, this second information can also be referred to as channel environment prediction related information, and it may include one or more of the following:

[0125] Interference information;

[0126] Channel matrix information;

[0127] Access network equipment computing resources;

[0128] Terminal computing power resources;

[0129] Connect to resources;

[0130] Access network device configuration information;

[0131] Access network device hardware information;

[0132] Information related to the first model.

[0133] The interference information may include downlink interference information from neighboring cells, such as downlink interference data reported by the terminal when measuring CSI-IM and NZP-CSI-RS. For example, the terminal can obtain interference information by measuring the downlink interference and / or thermal interference (IOT) values ​​(which can also be understood as interference rise values) received in each Physical Resource Block (PRB) and / or Subband when CSI-IM is configured, and report the interference information to the access network device. When the access network device sends the second information to the management device, the second information may include the interference information.

[0134] The channel matrix information may include channel data obtained by the terminal through CSI-RS channel estimation. For example, suppose that the p-th antenna of the access network device transmits CSI-RS (specifically downlink CSI-RS) on the k-th subcarrier, and the r-th antenna of the terminal receives the CSI-RS; at this time, the terminal can obtain the channel matrix H(k,r,p) through channel estimation and send the channel matrix H(k,r,p) to the access network device; when the access network device sends the second information to the management device, the second information may include the channel matrix H(k,r,p); or, after obtaining the channel matrix H(k,r,p), the terminal can use structural transformation methods such as multidimensional discrete Fourier transform (DFT) and / or singular value decomposition (SVD) to obtain the channel feature matrix, which can be specifically expressed as formula (2):

[0135]

[0136] in, Let G represent the channel feature matrix, and G represent the structural transformation matrix (such as the DFT matrix, orthogonal transform basis, etc.). The terminal can convert the channel feature matrix... The information is sent to the access network device; when the access network device sends the second information to the management device, the second information may include the channel feature matrix.

[0137] In practical applications, the access network device can send configuration information to the terminal via Radio Resource Control (RRC) connection (i.e., send configuration information to the terminal via RRC signaling). The configuration information may include CSI resource configuration and / or report configuration information. After receiving the configuration information, the terminal can obtain a measurement report based on the configured CSI resource measurement reference signal according to the configuration information, and report the measurement report according to the configuration information. The measurement report may include downlink interference information, channel matrix information, terminal computing resources, etc.

[0138] In practical applications, the management device can request (or instruct) the access network devices to report the second information. The access network devices can use the information contained in the measurement report, as well as their local information (such as resource information and hardware information), as the second information and send it to the management device. In other words, in one embodiment, the management device sends a fourth message to the plurality of access network devices, the fourth message being used to request the reporting of the second information; and receives the second information reported by the plurality of access network devices.

[0139] Of course, the management device can also subscribe to the second information of the access network devices. The access network devices report the second information to the management device according to the subscription period, or the access network devices report the second information to the management device when the subscription conditions are triggered. That is, in one embodiment, the management device subscribes to the second information of the multiple access network devices.

[0140] In practical applications, after receiving the second information from multiple access network devices, the management device can divide the multiple access network devices into M groups based on a similarity algorithm (which can also be understood as clustering multiple access network devices).

[0141] Specifically, in one embodiment, dividing the plurality of access network devices into M groups using the acquired second information includes:

[0142] For each of the plurality of second pieces of information, feature extraction is performed on the second piece of information to obtain the feature information of the access network device;

[0143] Based on the similarity of feature information of multiple access network devices, the multiple access network devices are divided into M groups.

[0144] In practical applications, the management device can determine the characteristic information corresponding to each access network device based on the second information of that access network device; wherein, the characteristic information may specifically include one or more of the following:

[0145] Hardware information of access network equipment, such as the type of access network equipment and antenna information of access network equipment;

[0146] Configuration information of access network equipment, such as time and frequency resource scheduling configuration parameters of access network equipment, and multiple-input multiple-output (MIMO) related configuration parameters of terminals (which can also be understood as users);

[0147] Terminal information (which can also be understood as user information) provided by access network equipment, such as terminal quality of service (QoS) requirements, terminal distribution, terminal mobility speed, terminal capabilities, and terminal service types;

[0148] Access network equipment's wireless channel environment information, such as interference information and channel matrix;

[0149] The first model information of the access network device, such as the algorithm characteristics of the first model and the type of the first model.

[0150] In practical applications, the management device can use a similarity algorithm to determine (or calculate) the similarity of the feature information of the access network devices. This similarity algorithm may include a cosine similarity algorithm.

[0151] For example, suppose the management device determines the similarity of two access network devices for feature A, the two access network devices being access network device 1 and access network device 2, the time series of feature A of access network device 1 can be represented as x, and the time series of feature A of access network device 2 can be represented as y. In this case, the similarity can be specifically expressed as formula (3).

[0152]

[0153] Where T(x,y) represents the similarity value between access network device 1 and access network device 2 for feature A, x i Let y represent the value of feature A at time i in x. i Let represent the value of feature A at time i in y. Here, a larger similarity value indicates that the features A of access network device 1 and access network device 2 are more similar, and they are more likely to be classified into the same group.

[0154] In practical applications, the management device can first determine one or more features from all features corresponding to the feature information, then determine the similarity between multiple access network devices for the determined one or more features, and based on preset grouping conditions, such as similarity greater than a preset threshold, divide the access network devices with high similarity into the same group, resulting in multiple groups. If the number of groups obtained reaches M, the multiple groups obtained can be used as the M groups; if the number of groups obtained is less than M, the management device can redetermine one or more features (i.e., features that are different from the previously determined features), then determine the similarity between multiple access network devices for the redetermined one or more features, and based on preset grouping conditions, further group the access network devices on the basis of the previously obtained multiple groups, until the number of groups obtained reaches M, and the multiple groups obtained are used as the M groups. Each of the M groups contains multiple access network devices. The value of M can be set according to actual needs. The larger the value of M, the more groups there are, the higher the similarity of the feature information of the multiple access network devices in each group, the higher the accuracy of the model obtained by federated learning for that group, and the worse the generalization (i.e., the poorer the accuracy for unknown datasets (i.e., the datasets corresponding to other groups)). Of course, the smaller the value of M, the fewer groups there are, the lower the similarity of the feature information of the access network devices in each group, the lower the accuracy of the model obtained by federated learning for that group, and the higher the generalization.

[0155] In practical applications, after dividing multiple access network devices into M groups, since the characteristic attributes of the multiple access network devices in each group are quite similar, and the data samples corresponding to each access network device exist independently, a horizontal federated learning framework can be adopted to perform federated learning on the multiple access network devices in the group. That is, the management device can determine multiple training devices and one aggregation device related to federated learning for each group.

[0156] When determining multiple training devices related to federated learning for a group, the management device can utilize information such as the similarity of local data of the access network devices included in the group and the computing resources of the access network devices to select multiple access network devices as training devices for federated learning. That is, when the second information includes relevant information about the local data of the access network devices, in one embodiment, the management device can determine multiple training devices for each group by utilizing the computing resource information and / or relevant information about the local data of the multiple access network devices corresponding to the group.

[0157] For example, the management device can first select multiple access network devices from the multiple access network devices included in the group that have computing power resources that meet the requirements of federated learning training (which can also be understood as meeting the computing power requirements of AI intelligent algorithms); then, the management device can further select multiple access network devices as the training devices from the selected multiple access network devices based on the similarity of local data (which can also be understood as the similarity of input features), and the local data of the selected multiple access network devices can represent all the local data of the multiple access network devices included in the group, for example, the proportion of the local data of the selected multiple access network devices to all local data meets a preset threshold.

[0158] When determining an aggregation device related to federated learning for a group, the management device may select a device whose computing resources and / or transmission latency meet the requirements of federated learning from among the multiple access network devices included in the group, the management device itself, and other management devices as the aggregation device.

[0159] Specifically, in one embodiment, determining an aggregation device for each group includes:

[0160] For each group, the aggregation device is determined using the computing resource information and / or transmission latency information related to multiple access network devices, management devices, and one or more other management devices besides the management device corresponding to the group. The one or more other management devices besides the management device may include OAM devices, Near-RTRIC devices, or Non-RT RIC devices, etc., and this embodiment of the application does not limit this; hereinafter, the multiple access network devices corresponding to the group, the management device, and one or more other management devices besides the management device are collectively referred to as the candidate aggregation devices corresponding to the group.

[0161] In practical applications, the management device can first select multiple candidate aggregation devices from the corresponding candidate aggregation devices in the group whose computing power resources meet the computing power requirements of federated learning (such as the computing power requirements related to aggregating and updating the model). Then, the management device can select the candidate aggregation device from the selected multiple candidate aggregation devices that meets the requirements of federated learning in terms of transmission latency and has the optimal overall latency as the aggregation device. The transmission latency includes the latency of the aggregation model and the latency of sending the aggregated model parameters to the multiple access network devices included in the group.

[0162] As can be seen from the above description, for each group, the management device can determine the aggregation device and training device related to federated learning. Thus, federated learning can be performed using the aggregation device and training device determined for each group to obtain the second model. Specifically, the management device can simultaneously determine the aggregation device and multiple training devices; or, the management device can first determine the aggregation device and then determine the multiple training devices; or, the management device can first determine the multiple training devices and then determine the aggregation device. This embodiment of the application does not limit this approach.

[0163] In practical applications, the second model can contain a single model (i.e., a model that can independently predict the channel environment) or multiple sub-models (i.e., multiple models that predict the channel environment through cooperation). When the second model contains a single model, the aggregation device, when executing the above process, can divide multiple access network devices into M groups based on the characteristics of the single model, and determine one aggregation device and multiple training devices for each group. At the same time, when the second model contains multiple sub-models, such as sub-models for predicting interference information, sub-models for predicting the channel matrix, etc., the management device can, for each sub-model, divide multiple access network devices into M groups based on the input and output data information of each access network device corresponding to the sub-model, and determine one aggregation device and multiple training devices for each group. In this case, the determined aggregation device and multiple training devices can perform federated learning to train the sub-model corresponding to the group.

[0164] Specifically, when the second model contains multiple sub-models, when the management device determines the aggregation device, it can consider the synchronization of each training device obtaining different sub-models from the aggregation device corresponding to different sub-models, so as to ensure that the parameters of multiple sub-models can be updated synchronously.

[0165] For example, suppose the second model contains two sub-models, namely sub-model 1 and sub-model 2, and access network device A belongs to multiple access network devices managed by the management device. When the management device groups sub-model 1, it assigns access network device A to group X, and when it groups sub-model 2, it assigns access network device A to group Y. At this time, when the management device determines the aggregation device 1 corresponding to sub-model 1 for group X and the aggregation device 2 corresponding to sub-model 2 for group Y, it can consider the time when aggregation device 1 sends the aggregated sub-model 1 parameters to access network device A, and the time when aggregation device 2 sends the aggregated sub-model 2 parameters to access network device A, and select appropriate aggregation devices 1 and 2 to make the two times as close as possible. In this way, access network device A can obtain the parameters of sub-model 1 and sub-model 2 as synchronously as possible, and realize the synchronous update of the two sub-models.

[0166] In practical applications, after determining multiple training devices and one aggregation device corresponding to a group, the management device can send instruction information to the aggregation device corresponding to that group. The instruction information is used to instruct the aggregation device to perform federated learning using multiple training devices to obtain a second model (i.e., the trained second model) corresponding to that group. Specifically, the instruction information may include: an initial model, relevant information of the multiple training devices corresponding to the group, and relevant information of the first access network devices included in the group. Here, the initial model can also be understood as an AI model to be trained for predicting the channel environment. The initial model may include a model determined by the management device based on the characteristics of the access network devices included in the group.

[0167] Upon receiving the instruction information sent by the management device, the aggregation device can determine multiple training devices based on the instruction information, and use the multiple training devices to perform federated learning to obtain the second model. The specific process (which can also be understood as the lifecycle process of the federated learning model) may include the following steps:

[0168] Step 1: Obtain the initial model using multiple training devices;

[0169] Specifically, the plurality of training devices may obtain the initial model from the aggregation device, or the plurality of training devices may obtain the initial model from other devices (such as the management device), and this application embodiment does not limit this.

[0170] Step 2: Each training device trains the initial model using local data (which can also be understood as improving the model or training the model locally) to obtain the trained model;

[0171] Step 3: The training device determines the improvement of the trained model compared to the initial model, and summarizes the improvement (which can also be understood as model summarization) to obtain updated information;

[0172] Step 4: The training device encrypts the update information to obtain encrypted update information, and then reports the encrypted update information to the aggregation device;

[0173] Step 5: The aggregation device uses the encrypted update information reported by multiple training devices to update the initial model, thereby obtaining the updated initial model.

[0174] In practical applications, steps 1 to 5 can be repeated to perform multiple rounds of training and improve model performance. The aggregation device can stop repeating the process when the number of repetitions reaches a preset threshold and / or the performance of the updated initial model meets preset requirements, and use the last updated initial model as the second model after training (also known as a shared model).

[0175] After determining the second model, the aggregation device can use the indication information to determine the first access network devices included in the group, and provide the trained second model to each of the first access network devices included in the group; accordingly, each first access network device can obtain the trained second model.

[0176] In practical applications, the management device can inform all access network devices in the group of the information of the aggregation device (such as identification information, address information, etc.), so that when the access network devices in the group determine to obtain the second model (i.e., determine to obtain the trained AI model using the second method mentioned above), they can directly obtain the second model from the aggregation device without triggering the management device to perform the federated learning process.

[0177] After the first access network device obtains the second model, it can send the second model to the terminal so that the terminal can use the second model to predict the channel environment.

[0178] Specifically, the first access network device can directly send the second model to the terminal. After receiving the second model, the terminal can directly use the second model to predict the channel environment. Alternatively, the first access network device can also prune and / or compress the second model according to the terminal's capabilities (such as computing resources) to obtain a processed second model, and then send the processed second model to the terminal. The terminal uses the processed second model to predict the channel environment. Or, the first access network device can directly send the second model to the terminal. After receiving the second model, the terminal can prune and / or compress the second model according to its own capabilities to obtain a processed second model, and then use the processed second model to predict the channel environment. The model used by the terminal to predict the channel environment is referred to as the terminal model. Furthermore, the process of the first access network device or the terminal using the second model to determine the terminal model, and the terminal using the terminal model to predict the channel environment, can be referred to as the access network device using the second model to predict the channel environment.

[0179] For example, after the terminal uses the terminal model to predict the channel environment, it can evaluate the prediction accuracy of the terminal model using the actual channel environment and the prediction results. When the prediction accuracy of the terminal model is lower than a preset threshold, the terminal can use locally collected data to fine-tune (or optimize) the terminal model to obtain a fine-tuned terminal model, and use the fine-tuned terminal model to predict the channel environment. Simultaneously, the terminal can report the model parameters of the fine-tuned terminal model to the first access network device. The first access network device can use the model parameters reported by the terminal to update the parameters of the second model, obtaining updated second model parameters, and report the updated second model parameters to the aggregation device. The aggregation device can use the updated second model parameters reported by the first access network device to update the second model (which can also be understood as performing a model aggregation process).

[0180] In practical applications, the management device can obtain the performance indicators of the one or more first access network devices using the second model to predict the channel environment, and if the performance indicators indicate that the performance of the second model is poor (i.e. the predicted channel environment cannot optimize network performance), the second model can be retrained to ensure the performance of the channel environment prediction.

[0181] Based on this, in one embodiment, the method may further include:

[0182] The fifth piece of information is determined, which indicates that one or more second access network devices are unable to optimize network performance using the channel environment predicted by the second model;

[0183] The second model is retrained to obtain a retrained second model; a retrained second model is provided for each of the one or more second access network devices.

[0184] In practical applications, the second access network device includes the access network device among the one or more first access network devices that, according to the channel environment prediction obtained using the second model, cannot optimize network performance.

[0185] The management device can obtain the prediction performance of the access network devices in predicting the channel environment using the second model, and determine which access network devices belong to the second access network devices based on the obtained prediction performance. If the number of second access network devices reaches a preset threshold, it determines that the second model needs to be retrained, i.e., it determines the fifth information. Specifically, in one embodiment, the management device obtains the prediction performance of the multiple access network devices in predicting the channel environment using the second model; and identifies access network devices with prediction performance lower than a first threshold as the second access network devices. The first threshold can be set according to actual needs, and this embodiment does not limit it. The prediction performance may include the performance indicators of the second model and / or the performance indicators of the AMC algorithm. The performance indicators of the second model may include one or more of the following:

[0186] Training model accuracy;

[0187] Data distribution coverage;

[0188] Training convergence time;

[0189] Generalization;

[0190] robustness;

[0191] Energy consumption.

[0192] The performance metrics of the AMC algorithm may include one or more of the following:

[0193] Throughput;

[0194] Communication reliability;

[0195] Delay;

[0196] IBLER.

[0197] In practical applications, the management device can obtain the predicted performance of the access network device in predicting the channel environment using the second model from the access network device and / or related aggregation device that use the second model for performance prediction through a subscription method.

[0198] Of course, the access network device using the second model for performance prediction can also use local relevant information (such as training accuracy, data range, energy consumption, etc.) and / or relevant information obtained from the corresponding aggregation device (such as generalization, robustness, etc.) to determine its own prediction performance for predicting the channel environment using the second model, and determine that it belongs to the second access network device if the prediction performance is lower than a preset threshold, and inform the management device. That is, in one embodiment, the management device receives a sixth piece of information sent by one or more access network devices, the sixth piece of information indicating that the prediction performance for predicting the channel environment using the second model is lower than a second threshold; the access network device corresponding to the sixth piece of information is designated as the second access network device. In this way, the management device can determine that the second model needs to be retrained when the number of second access network devices reaches a preset threshold, that is, determine the fifth piece of information.

[0199] After determining the fifth piece of information, the management device can retrain the second model for each group using a pre-determined aggregation device and multiple training devices. In other words, in one embodiment, the management device reuses the aggregation device and multiple training devices to train the second model.

[0200] Of course, the management device can also regroup the managed access network devices, and then, for each regrouped group, determine an aggregation device and multiple training devices, and use the determined aggregation device and multiple training devices to train the second model. That is, in one embodiment, the management device uses the acquired second information to regroup the multiple access network devices into N groups, where N is an integer greater than or equal to 1; for each regrouped group, it determines an aggregation device and multiple training devices; and uses the aggregation device and multiple training devices to train the second model.

[0201] In practical applications, the management device can determine whether the second model is overfitting or underfitting based on its performance metrics. If the second model is overfitting, it indicates that the training dataset is not rich enough. When regrouping multiple access network devices into N groups based on similarity, the management device can lower the similarity threshold to include more access network devices in each group, thus enriching the training dataset. In this case, the number of groups is smaller; that is, in one embodiment, N is less than M when the second model is overfitting. Conversely, if the second model is underfitting, it indicates that the complexity of the second model is insufficient to represent the information in the training dataset. When regrouping multiple access network devices into N groups based on similarity, the management device can increase the similarity threshold to include fewer access network devices in each group, reducing the data diversity of access network devices in each group. In this case, the number of groups is larger; that is, in one embodiment, N is greater than M when the second model is underfitting.

[0202] The channel environment prediction method provided in this application embodiment includes a management device determining first information, wherein the management device is used to manage at least multiple access network devices, the first information indicating that a local first model of one or more first access network devices cannot be trained and / or the channel environment predicted by the local first model cannot optimize network performance, the first model being used to predict the channel environment; and providing a corresponding second model for each of the one or more first access network devices, the second model being used to predict the channel environment, the second model being obtained by federated learning using the multiple access network devices managed by the management device. The solution provided in this application embodiment involves an access network device first using a local model to predict the channel environment. When a management device that manages multiple access network devices determines that one or more access network devices have difficulty training a local model and / or the performance of the local model is insufficient to meet requirements (e.g., the predicted channel environment cannot optimize network performance), the management device provides these access network devices with a model obtained through federated learning from multiple access network devices. In this way, on the one hand, the management device can fully utilize the data and computing resources of multiple access network devices to train the model through federated learning, thereby obtaining a channel environment prediction model with better performance and lower resource requirements for each access network device. On the other hand, after receiving the model provided by the management device, the access network device can use the provided model to predict the channel environment to achieve better channel environment prediction performance (e.g., accuracy) and ensure that the predicted channel environment can meet the requirements for optimizing network performance.

[0203] The following section provides a more detailed description of this application with reference to application examples.

[0204] In a communication system, when a base station (i.e., the aforementioned access network equipment) initiates data service transmission with a terminal, the base station uses the CQI most recently reported by the terminal to determine the channel quality between the base station and the terminal. However, there is a time delay between the time when the terminal determines the CQI and the time when it initiates data service transmission. The CQI previously reported by the terminal may be outdated for the time when the data service transmission is about to be initiated. This will cause the base station to be unable to accurately determine the channel quality at the time of data service transmission based on the CQI, and thus cannot select an MCS suitable for the channel quality at that time.

[0205] Based on this, this application proposes an MCS selection scheme, in which the base station selects the MCS using the predicted CQI of the future data service transmission time. That is, the terminal predicts the CQI of the future data service transmission time and reports the predicted CQI to the base station. Specifically, the terminal can predict the CQI of the future data service transmission time based on uplink interference values ​​and channel matrix prediction values.

[0206] In practical applications, base stations can train a model for predicting the channel matrix (hereinafter referred to as the channel matrix model) based on the historical channel matrix data reported by the terminal.

[0207] The process of constructing a channel matrix model may include: Figure 2 As shown, the base station receives historical channel matrix data reported by the terminal and performs analysis on data within a certain period of time (e.g., ...). Figure 2 The channel matrix data received during the time interval from time Tm to time T is preprocessed to obtain the input data for the channel matrix model; simultaneously, the base station can determine the prediction step size (e.g., based on channel environment strategies) Figure 2 The prediction step size is k), and the channel matrix data at the time corresponding to the prediction step size is used as the output data of the channel matrix model. Based on the input and output data, a channel matrix model is constructed. The constructed channel matrix model is then trained using the input data to obtain the trained channel matrix model. The historical channel matrix data reported by the terminal can include the time series corresponding to each subcarrier j. For example, for the channel between the r-th antenna of the base station and the p-th antenna of the terminal, the CSI-RS channel matrix data of the j1-th subcarrier from time Tm to time T can be expressed as {H... j1_T-m (r,p),H j1_T-m+1 (r,p),…,H j1_T (r,p)}, at this time, the output data of the channel matrix model (which can also be understood as the prediction result) can be expressed as H T+k (j,r,p).

[0208] After training the channel matrix model, the base station can train a model for predicting uplink interference values ​​(hereinafter referred to as the uplink interference model) based on the historical uplink interference data reported by the terminal.

[0209] The process of building an uplink interference model may include: the base station receiving historical uplink interference data reported by the terminal, and analyzing the data over a period of time (e.g., ...). Figure 2 The uplink interference data received during the time interval from time Tm to time T is preprocessed to obtain the input data for the uplink interference model. Simultaneously, the base station can determine the prediction step size based on the channel environment strategy, and use the uplink interference data at the time corresponding to the prediction step size as the output data for the uplink interference model. Thus, based on the input and output data, an uplink interference model is constructed. The constructed uplink interference model is then trained using the input data to obtain the trained uplink interference model. The historical uplink interference data reported by the terminal can include the uplink interference values ​​(which can be expressed as interference) received by the terminal within a preset frequency band (or reporting frequency band) for each subcarrier, RB, or RBG. For example, the time series of uplink interference for the j1-th subcarrier during the time interval from time Tm to time T can be expressed as: {I j1_T-m ,I j1_T-m+1 ,…,I j1_T At this point, the output data of the uplink interference model (which can also be understood as the prediction result) can be expressed as I. T+k .

[0210] In practical applications, the terminal can use the uplink interference model to predict the uplink interference prediction value at the subcarrier / RB / RBG level at the time of future data service transmission, and use the method of averaging the values ​​at the subcarrier / RB / RBG level to convert the uplink interference prediction value at the subcarrier / RB / RBG level into the average uplink interference value for each subcarrier / RB / RBG on the frequency band used by the terminal.

[0211] Specifically, the terminal can use the channel matrix model to predict the channel matrix at time T+k (which can also be understood as tti at time T+k), and obtain the predicted channel matrix value at time T+k, which can be expressed as H. T+k (j,r,p); then, the terminal can utilize structural transformation methods such as multidimensional DFT and / or SVD to transform H. T+k By performing structural transformations on (j,r,p), the predicted channel feature matrix at time T+k is obtained. The terminal can utilize The predicted value of the signal strength (or useful signal strength) of each subcarrier received by the terminal can be determined by formula (4):

[0212]

[0213] Among them, X T+k (j,p) represents the reference signal sent by the user at time T+k, S T+k (j,r) represents the terminal's predicted signal strength for the j-th subcarrier at time T+k, and P represents the total number of antennas of the terminal (which can also be understood as the total number of antennas of the transmitting end); here, the terminal can determine X based on the base station's configuration information for the reference signal. T+k (j,p).

[0214] Determine the S T+k After (j,r), the terminal can use the terminal's predicted signal strength for each subcarrier at time T+k to determine the predicted average signal strength (or average useful signal strength) of the scheduling user for each subcarrier / RB / RBG in a certain frequency domain at time T+k. The specific calculation formula can be expressed as formula (5):

[0215]

[0216] Where J represents the total number of subcarriers for the scheduled user in a certain frequency domain, R represents the total number of antennas of the base station (which can also be understood as the number of receiving antennas), N_sc represents the number of subcarriers contained in each subcarrier / RB / RBG, and S T+k This represents the predicted average signal strength of each subcarrier / RB / RBG in a certain frequency domain at time T+k by the scheduling user.

[0217] Determine the S T+k Then, the terminal can utilize S T+k and I T+k The channel signal-to-noise ratio (SNR) at time T+k (which can also be understood as uplink SNR or channel quality) can be determined by formula (6):

[0218]

[0219] in, SINR represents the intensity of Gaussian white noise. T+k This represents the channel signal-to-noise ratio (SNR) at time T+k. The terminal can utilize this SINR. T+k Determine the CQI at time T+k. T+k (The predicted CQI for future data service transmission times); then, the terminal can use the CQI T+k The data is sent to the base station, which can then utilize CQI. T+k Select MCS.

[0220] In practical applications, when base stations train the aforementioned channel matrix model and uplink interference model, insufficient computing power and / or data may lead to performance issues that fail to meet requirements. In such cases, federated learning can be used with multiple base stations to train the channel matrix model and uplink interference model, and the trained model can then be deployed on the base stations requiring channel environment prediction. This scheme of using federated learning to train the channel matrix model and uplink interference model can also be called a federated learning-based channel environment prediction scheme.

[0221] However, in practical applications, channel environment prediction schemes based on federated learning may have the following drawbacks:

[0222] 1) The channel environment prediction scheme based on federated learning needs to train the model on multiple base stations, and the computational cost is much higher than that of the traditional measurement-based channel environment algorithm. Not all situations require the channel environment prediction scheme and federated learning model framework. It is necessary to design how to trigger the channel environment prediction scheme and federated learning model framework.

[0223] 2) Federated learning relies on the participation of multiple base stations. The data distribution of each base station is quite different, which leads to a decrease in the accuracy of the aggregated model. It is necessary to design a model based on the similarity of data distribution to group base stations, and base stations in the same group train a model together.

[0224] 3) Among base stations in the same group, each base station must be trained independently on similar data, which will lead to redundant computation and waste a lot of time and resources. Strategies need to be designed to coordinate the training process between different base stations.

[0225] 4) During the federated learning process, each base station needs to use the connection resources between the aggregation node and the device to report model parameters and receive model update parameters. If the connection resources cannot meet the latency requirements of the base station's model parameter reporting and updating, it will affect the performance of the model. Therefore, the federated learning model aggregation node needs to be selected based on the aggregation capability and the connection resources with other base stations.

[0226] To avoid the aforementioned drawbacks, this application proposes a channel environment prediction method based on federated learning in its application examples, such as... Figure 3 As shown, it includes the following steps:

[0227] Step 301: Trigger the channel environment prediction scheme based on federated learning;

[0228] Specifically, this may include the following steps:

[0229] Step 3011: The intelligent network device (i.e., the aforementioned management device, such as RIC) sends a subscription request to the base station (i.e., the aforementioned access network device, such as RAN) within its service range to request subscription to network performance indicators such as IBERL, throughput, and latency of the base station and / or terminal (such as UE) (which can also be understood as the system monitoring base station-related performance indicators); at the same time, the intelligent network device can also send the method for triggering the reporting of subscription information (which can also be understood as the triggering condition, which may specifically include a threshold) to the base station;

[0230] Step 3012: The base station detects network performance indicators such as IBERL, throughput, and latency, and determines whether the network performance indicators meet the triggering conditions (or can be understood as not meeting the threshold).

[0231] Step 3013: When the network performance indicators meet the triggering conditions, the base station sends the base station identifier (such as base station ID) and network performance indicators to the intelligent network device.

[0232] Step 3014: After receiving the information reported by the base station, the intelligent network device sends an instruction message to the base station, instructing the base station to trigger the intelligent resource optimization algorithm scheme (which can also be understood as triggering the channel environment prediction scheme), that is, to perform channel environment prediction based on the locally trained model (i.e. the first model mentioned above), thereby optimizing the MCS selection, which can also be understood as performing code domain resource optimization.

[0233] In practical applications, the mechanism provided in steps 3011 to 3014 can also be understood as a mechanism for triggering resource scheduling optimization schemes based on wireless environment prediction.

[0234] When a base station receives an instruction and trains a channel environment prediction model according to the intelligent resource optimization algorithm, if the base station meets one or more of the following conditions, it will be difficult for the base station to obtain a channel environment prediction model that meets the accuracy requirements through local training, and it will need to obtain a model trained through federated learning (i.e., the second model mentioned above):

[0235] The computing power of the base station is insufficient to support AI model training;

[0236] The energy-saving conditions of the base station (which can also be understood as energy consumption or electricity consumption) cannot support AI model training;

[0237] The accuracy of the trained AI model cannot reach the threshold for deployment.

[0238] The convergence time during AI model training does not meet the user's QoS requirements.

[0239] In practical applications, if the intelligent network device has already trained a model for channel environment prediction using multiple base stations through federated learning, the base station that meets one or more of the above conditions can directly obtain the model based on federated learning from the related equipment. If the intelligent network device has not previously trained a model for channel environment prediction using multiple base stations through federated learning, then step 3015a or step 3015b is executed to perform federated learning and obtain a model for channel environment prediction.

[0240] Step 3015a: The base station sends triggering information to the intelligent network device to trigger the federated learning-based channel environment prediction scheme;

[0241] Step 3015b: The intelligent network device detects multiple base stations under its management. If the number of base stations that meet the above conditions is greater than a preset threshold, the channel environment prediction scheme based on federated learning is triggered.

[0242] In practical applications, after triggering the channel environment prediction scheme based on federated learning, the intelligent network device can execute steps 302 to 303 to obtain the channel environment prediction model through federated learning.

[0243] Step 302: Federated learning model orchestration management (which can also be understood as network element collaborative orchestration management based on federated learning training of channel environment prediction models);

[0244] Specifically, this may include the following steps:

[0245] Step 3021: The intelligent network device sends subscription requests to multiple base stations under its management to request information related to the base stations, such as interference information, channel matrix information, computing resources, connection resources, and locally trained model information (e.g., the input and output feature dimensions and model expressions of the intelligent algorithm model (i.e., the locally trained model mentioned above)).

[0246] Step 3022: The base station reports the above information according to the subscription request;

[0247] Step 3023: The intelligent network device groups the base stations based on the similarity algorithm (which can also be understood as performing base station clustering) to obtain multiple groups (i.e., the above M groups), and then executes steps 3024 and 3025 for each group;

[0248] Specifically, intelligent network devices group the multiple base stations they manage based on base station-related interference information, channel matrix information, computing resources, and connection resources. Simultaneously, intelligent network devices can further aggregate and classify base stations based on the input-output feature dimensions and model expressions of the intelligent algorithm model, enabling multiple base stations within each group (or, in other words, base stations within the same group) to share the intelligent algorithm model (i.e., to use the same model trained using federated learning).

[0249] Here, the intelligent algorithm model can include an uplink interference time series prediction model using the LSTM algorithm (i.e., the uplink interference model mentioned above) and a channel matrix time series prediction model (i.e., the channel matrix model mentioned above). The input to the uplink interference time series prediction model can include the measured received interference power (RIP), and the output can include the predicted uplink interference values ​​at future times. The input to the channel matrix time series prediction model can include the channel matrix H(j,r,p), and the output can include the channel matrix at future times.

[0250] In practical applications, since the input-output logic of the uplink interference time series prediction model and the channel matrix time series prediction model is universal across different base stations, base stations can be grouped based on a similarity algorithm, taking into account the input-output logic of each base station's local uplink interference time series prediction model. After grouping, the multiple base stations in each group can collaboratively train a single uplink interference time series prediction model through federated learning. Similarly, base stations can be grouped based on a similarity algorithm, taking into account the input-output logic of each base station's local channel matrix time series prediction model. After grouping, the multiple base stations in each group can collaboratively train a single channel matrix time series prediction model through federated learning.

[0251] In practical applications, the intelligent network device can group base stations for the uplink interference time series prediction model and the channel matrix time series prediction model respectively. That is, the base station grouping process of the federated learning framework of the above two models is independent of each other. Each base station can be divided into two different groups. The base stations in one group can be trained by federated learning to train the uplink interference time series prediction model, and the base stations in the other group can be trained by federated learning to train the channel matrix time series prediction model.

[0252] Specifically, since the time fluctuation period, trend, and amplitude of the input sequences (which can also be understood as RIP-related time sequences) of the uplink interference time series prediction models of different base stations are not completely the same, when the intelligent network device groups base stations, it can use a similarity algorithm to group base stations whose similarity between the input sequences of the uplink interference time series prediction models in the same historical time period is higher than a preset threshold.

[0253] Meanwhile, since the time fluctuation period, trend, and amplitude of the input channel matrix sequence of the local channel matrix time series prediction model of different base stations are not completely the same, when the intelligent network device groups base stations, it can use a similarity algorithm to group base stations whose input channel matrix sequence similarity is higher than a preset threshold in the same historical time period.

[0254] After the intelligent network device groups the base stations, if the data information of the related features of different base stations within each group is more similar (i.e., the similarity between different base stations is higher), the prediction accuracy of the model obtained through federated learning is higher (i.e., the prediction accuracy after aggregation is higher), but the generalization is poor. If the data information of the related features of different base stations within each group is more dissimilar (i.e., the similarity between different base stations is lower), the prediction accuracy of the model obtained through federated learning is lower (i.e., the prediction accuracy after aggregation is lower), but the generalization is strong.

[0255] Step 3024: For each group, the intelligent network device selects multiple local model training nodes (i.e., the training devices mentioned above) corresponding to that group;

[0256] In practical applications, intelligent network devices can select multiple local model training nodes from among the base stations within a group, based on data similarity between base stations and the computing resources of each base station. Specifically, one or more of the following principles can be followed:

[0257] The computing resources of the local model training node are sufficient to meet the computing power requirements of the AI ​​intelligent AMC algorithm;

[0258] Based on the data similarity of the input features, the data information of the selected multiple local model training nodes can represent a certain percentage threshold of the data information corresponding to all base stations in the group.

[0259] Step 3025: For each group, the intelligent network device selects the global model aggregation node corresponding to that group (i.e., the aggregation device mentioned above, which can also be understood as the aggregation node);

[0260] In practical applications, intelligent network devices can select global model aggregation nodes based on connection resources and computing power resources with local training nodes. Specifically, the following principles can be followed:

[0261] The computing resources of the global model aggregation node meet the computing power requirements for aggregating and updating AI models;

[0262] The connection resources of the global model aggregation node meet the latency requirements for aggregating and distributing AI model parameters to each local model training node.

[0263] For each base station, the transmission delay from the base station to the global model aggregation node corresponding to the uplink interference time series prediction model is approximately the same as the transmission delay from the base station to the global model aggregation node corresponding to the channel matrix time series prediction model. This ensures the synchronization of different model parameters sent by different global model aggregation nodes to the base station.

[0264] Step 303: Lifecycle management of federated learning models;

[0265] Specifically, for each model (such as an uplink interference time series prediction model or a channel matrix time series prediction model), the intelligent network device sends an instruction to the global model aggregation node in each group based on that model. This instruction instructs the global model aggregation node to perform federated learning using multiple local training nodes corresponding to that group, thereby obtaining the trained model. Then, the global model aggregation node can provide the model parameters of the trained model to the base station that needs to obtain the federated learning model. Correspondingly, the base station receives the model parameters of the federated learning model (i.e., the model parameters that have been trained to achieve the required deployment accuracy). This process can also be referred to as the lifecycle process of the federated learning model.

[0266] After obtaining the model parameters provided by the global model aggregation node, the base station determines the channel environment prediction model (hereinafter referred to as the federated learning model). The base station can compress and / or prune the federated learning model according to the terminal's capabilities to obtain the processed model, and then send the processed model to the terminal. Alternatively, the base station can directly send the federated learning model to the terminal, and the terminal can compress and / or prune the federated learning model (which can also be understood as fine-tuning).

[0267] Step 304: Performance evaluation and detection of intelligent AMC algorithm based on federated learning (which can also be understood as performance evaluation of intelligent AMC algorithm).

[0268] In practical applications, after the base station obtains the federated learning model, it may still experience poor performance when predicting the channel environment. In this case, the base station can evaluate the performance of the federated learning model and the AMC algorithm, obtain performance evaluation results, and send these results to the intelligent network device. The intelligent network device can then retrain the federated learning model based on the performance evaluation results.

[0269] Based on this, step 304 may specifically include:

[0270] Step 3041: The base station sends the performance evaluation results to the intelligent network device;

[0271] The performance evaluation results can specifically include performance metrics of the federated learning model and performance metrics of the intelligent AMC algorithm. The performance metrics of the federated learning model can include: prediction accuracy, data distribution coverage, convergence time during training, generalization ability, robustness, and energy consumption; the performance metrics of the intelligent AMC algorithm can specifically include: throughput, communication reliability, latency, and IBLER.

[0272] Step 3042: The intelligent network device manages the channel environment prediction scheme adopted by the base station based on the performance index evaluation results.

[0273] Specifically, intelligent network devices can activate relevant models in base stations and instruct them to stop predicting network environments when performance evaluation results indicate that the network change rate of the base station is not significant and network environment prediction is not required, thereby saving computing resources. Alternatively, intelligent network devices can instruct the use of locally trained models for network environment prediction when performance evaluation results indicate changes in the local performance of the base station (such as expansion of computing resources), without the need for additional federated learning, thus saving computing resources. Or, intelligent network devices can regroup the base stations they manage and re-perform federated learning when performance evaluation results indicate that the current federated learning model is not performing well, thereby obtaining a federated learning model with better performance.

[0274] When the managed base stations are regrouped, the intelligent network device can dynamically adjust the similarity threshold based on performance evaluation results, and adjust the number of groups and the number of base stations in each group according to the similarity threshold, in order to avoid overfitting or underfitting of the model and improve the model's performance. Specifically,

[0275] When the performance evaluation results indicate that the current federated learning model is overfitting, the intelligent network device can determine that the dataset contained in each group of base stations is not rich enough. In this case, the similarity threshold can be lowered to include more base stations in each group, thereby enriching the dataset contained in each group of base stations and avoiding overfitting. At the same time, when the performance evaluation results indicate that the current federated learning model is underfitting, the intelligent network device can determine that the complexity of the federated learning model obtained by each group of base stations through federated learning is insufficient to express the dataset information contained in that group of base stations (i.e., the dataset is too scattered). In this case, the similarity threshold can be increased to include fewer base stations in each group, thereby reducing the diversity of the dataset contained in each group of base stations and avoiding underfitting.

[0276] This application proposes a channel environment prediction scheme based on a federated learning framework. It also designs triggering conditions for this scheme: a resource allocation scheme based on channel environment prediction is triggered when the performance indicators of traditional resource scheduling algorithms fail to meet user QoS requirements; and the scheme is triggered when the base station environment for channel environment prediction is insufficient to support AI / ML algorithm model training. For collaborative orchestration management of federated learning, base station grouping rules are designed. Base stations within a group can jointly train the same channel environment prediction model under the federated learning framework. For example, base stations are grouped based on the similarity of algorithm-related feature data distribution; local model training nodes are selected based on data similarity between base stations within the group and the computing resources of each base station; and global model aggregation nodes are selected based on connection resources and computing resources with the local training nodes. Furthermore, after federated learning, base station grouping, selection of local training nodes, and aggregation nodes can be dynamically adjusted based on the performance evaluation of AI / ML algorithms and intelligent algorithms.

[0277] The solution proposed in this application example has the following advantages:

[0278] 1) By using federated learning to enrich data resources with data from different base stations, the accuracy and generalization performance of the model can be improved.

[0279] 2) Using multiple base stations to train the model simultaneously greatly saves the time for training the model (which can also be understood as the time for building the model);

[0280] 3) For base stations that cannot meet the requirements for model training due to insufficient computing power, the base station grouping rules can be used to find the base station group in which they belong, obtain the applicable global model obtained through federated training (i.e. the federated learning model mentioned above), and use the global model for inference and prediction, which can save computing power and obtain the performance gains brought by intelligent algorithms.

[0281] To implement the method of the embodiments of this application, the embodiments of this application also provide a channel environment prediction device, which is disposed on a management device. The management device is used to manage at least multiple access network devices, such as... Figure 4 As shown, the device includes:

[0282] The determining unit 401 is used to determine first information, the first information representing that a first model local to one or more first access network devices cannot be trained and / or the channel environment predicted by the local first model cannot optimize network performance, the first model being used to predict the channel environment;

[0283] The providing unit 402 is used to provide a corresponding second model for each of the one or more first access network devices. The second model is used to predict the channel environment and is obtained by federated learning using multiple access network devices managed by the management device.

[0284] In one embodiment, the device may further include:

[0285] The acquisition unit is used to acquire second information of multiple access network devices, the second information including relevant information of the access network devices;

[0286] The determining unit 401 is further configured to:

[0287] Using the acquired second information, the multiple access network devices are divided into M groups, each group containing multiple access network devices, where M is an integer greater than or equal to 1;

[0288] For each group, an aggregation device and multiple training devices are determined. The training devices are used to train the second model using local data and obtain training results. The training devices belong to multiple access network devices corresponding to the group. The aggregation device is used to aggregate the training results obtained by the multiple training devices to obtain the trained second model and provide the trained second model to each first access network device included in the group. Federated learning is performed using the aggregation device and multiple training devices.

[0289] In one embodiment, the determining unit 401 is specifically used for:

[0290] For each of the plurality of second pieces of information, feature extraction is performed on the second piece of information to obtain the feature information of the access network device;

[0291] Based on the similarity of feature information of multiple access network devices, the multiple access network devices are divided into M groups.

[0292] In one embodiment, the second information includes information related to the local data of the access network device, and the determining unit 401 is specifically used for:

[0293] For each group, multiple training devices are determined by utilizing the computing power resource information and / or relevant information of the local data of the multiple access network devices corresponding to the group.

[0294] In one embodiment, the determining unit 401 is specifically used for:

[0295] For each group, the aggregation device is determined by utilizing the computing power resource information and / or transmission latency information related to multiple access network devices corresponding to the group, the management device, and one or more other management devices besides the management device.

[0296] In one embodiment, the acquisition unit is specifically used for:

[0297] Send a fourth message to the plurality of access network devices, the fourth message being used to request the reporting of the second message; receive the second message reported by the plurality of access network devices;

[0298] or,

[0299] Subscribe to the second information of the multiple access network devices.

[0300] In one embodiment, the determining unit 401 is further configured to:

[0301] The fifth piece of information is determined, which indicates that one or more second access network devices are unable to optimize network performance using the channel environment predicted by the second model;

[0302] The second model is retrained to obtain a retrained second model; a retrained second model is provided for each of the one or more second access network devices.

[0303] In one embodiment, the acquisition unit is further configured to:

[0304] Obtain the prediction performance of the plurality of access network devices in predicting the channel environment using the second model; and designate the access network devices with prediction performance lower than a first threshold as the second access network devices.

[0305] or,

[0306] Receive sixth information sent by one or more access network devices, the sixth information indicating that the prediction performance of the channel environment using the second model is lower than a second threshold; designate the access network device corresponding to the sixth information as the second access network device.

[0307] In one embodiment, the determining unit 401 is specifically used for:

[0308] The second model is trained by reusing the aggregation device and multiple training devices;

[0309] or,

[0310] Using the acquired second information, the multiple access network devices are re-divided into N groups, where N is an integer greater than or equal to 1; for each re-divided group, an aggregation device and multiple training devices are determined; the second model is trained using the aggregation device and multiple training devices.

[0311] In practical applications, the determining unit 401 can be implemented by the processor in the channel environment prediction device, and the providing unit 402 and the obtaining unit can be implemented by the processor in the channel environment prediction device in combination with the communication interface.

[0312] It should be noted that the channel environment prediction device provided in the above embodiments is only illustrated by the division of the above-described program units when performing channel environment prediction. In practical applications, the above processing can be assigned to different program units as needed, that is, the internal structure of the device can be divided into different program units to complete all or part of the processing described above. In addition, the channel environment prediction device and the channel environment prediction method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0313] Based on the hardware implementation of the above program modules, and in order to implement the method of the embodiments of this application, the embodiments of this application also provide a management device, which is used to manage at least multiple access network devices, such as... Figure 5 As shown, the management device 500 includes:

[0314] The communication interface 501 enables information exchange with other devices;

[0315] The processor 502 is connected to the communication interface 501 to enable information interaction with other devices and to execute the methods provided by one or more of the above-mentioned technical solutions when running a computer program;

[0316] The computer program is stored in memory 503.

[0317] Specifically, the processor 502 is used for:

[0318] The communication interface 501 determines first information, which indicates that a first model local to one or more first access network devices cannot be trained and / or the channel environment predicted by the local first model cannot optimize network performance. The first model is used to predict the channel environment. The communication interface 501 also provides a corresponding second model for each of the one or more first access network devices. The second model is used to predict the channel environment and is obtained by federated learning using multiple access network devices managed by the management device.

[0319] In one embodiment, the processor 502 is further configured to:

[0320] The second information of multiple access network devices is obtained through the communication interface 501, and the second information includes relevant information of the access network devices.

[0321] Using the acquired second information, the multiple access network devices are divided into M groups, each group containing multiple access network devices, where M is an integer greater than or equal to 1;

[0322] For each group, an aggregation device and multiple training devices are determined. The training devices are used to train the second model using local data and obtain training results. The training devices belong to multiple access network devices corresponding to the group. The aggregation device is used to aggregate the training results obtained by the multiple training devices to obtain the trained second model and provide the trained second model to each first access network device included in the group. Federated learning is performed using the aggregation device and multiple training devices.

[0323] In one embodiment, the processor 502 is specifically used for:

[0324] For each of the plurality of second pieces of information, feature extraction is performed on the second piece of information to obtain the feature information of the access network device;

[0325] Based on the similarity of feature information of multiple access network devices, the multiple access network devices are divided into M groups.

[0326] In one embodiment, the second information includes information related to local data of the access network device, and the processor 502 is specifically used for:

[0327] For each group, multiple training devices are determined by utilizing the computing power resource information and / or relevant information of the local data of the multiple access network devices corresponding to the group.

[0328] In one embodiment, the processor 502 is specifically used for:

[0329] For each group, the aggregation device is determined by utilizing the computing power resource information and / or transmission latency information related to multiple access network devices corresponding to the group, the management device, and one or more other management devices besides the management device.

[0330] In one embodiment, the processor 502 is specifically used for:

[0331] The fourth information is sent to the plurality of access network devices through the communication interface 501, the fourth information being used to request the reporting of the second information; and the second information reported by the plurality of access network devices is received.

[0332] or,

[0333] The communication interface 501 is used to subscribe to the second information of the plurality of access network devices.

[0334] In one embodiment, the processor 502 is further configured to:

[0335] The fifth piece of information is determined, which indicates that one or more second access network devices are unable to optimize network performance using the channel environment predicted by the second model;

[0336] The second model is retrained to obtain a retrained second model; a retrained second model is provided for each of the one or more second access network devices.

[0337] In one embodiment, the processor 502 is further configured to:

[0338] The prediction performance of the multiple access network devices in predicting the channel environment using the second model is obtained through the communication interface 501; the access network devices with prediction performance lower than a first threshold are designated as the second access network devices.

[0339] or,

[0340] The communication interface 501 receives a sixth message sent by one or more access network devices, the sixth message indicating that the prediction performance of the channel environment using the second model is lower than a second threshold; the access network device corresponding to the sixth message is designated as the second access network device.

[0341] In one embodiment, the processor 502 is specifically used for:

[0342] The second model is trained by reusing the aggregation device and multiple training devices;

[0343] or,

[0344] Using the acquired second information, the multiple access network devices are re-divided into N groups, where N is an integer greater than or equal to 1; for each re-divided group, an aggregation device and multiple training devices are determined; the second model is trained using the aggregation device and multiple training devices.

[0345] It should be noted that the specific processing procedures of the processor 502 and the communication interface 501 can be understood by referring to the above method.

[0346] Of course, in practical applications, the various components in the management device 500 are coupled together through the bus system 504. It can be understood that the bus system 504 is used to implement communication between these components. In addition to the data bus, the bus system 504 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 5 The general designated all buses as Bus System 504.

[0347] The memory 503 in this embodiment is used to store various types of data to support the operation of the management device 500. Examples of such data include any computer program used to operate on the management device 500.

[0348] The methods disclosed in the embodiments of this application can be applied to the processor 502, or implemented by the processor 502. The processor 502 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor 502 or by instructions in the form of software. The processor 502 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 502 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software modules may be located in a storage medium, which is located in the memory 503. The processor 502 reads the information in the memory 503 and combines its hardware to complete the steps of the aforementioned method.

[0349] In an exemplary embodiment, the management device 500 may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components to perform the aforementioned method.

[0350] It is understood that the memory (memory 503) in this embodiment of the application can be volatile memory or non-volatile memory, or both. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); the magnetic surface memory can be disk storage or magnetic tape storage. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM).The memories described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable types of memories.

[0351] In an exemplary embodiment, this application also provides a storage medium, namely a computer storage medium, specifically a computer-readable storage medium, such as a memory 503 storing a computer program, which can be executed by the processor 502 of the management device 500 to complete the steps described in the aforementioned method. The computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, Flash Memory, magnetic surface memory, optical disc, or CD-ROM.

[0352] In an exemplary embodiment, this application also provides a computer program product, including a computer program that can be executed by the processor 502 of the management device 500 to complete the steps described in the aforementioned method.

[0353] It should be noted that terms such as "first" and "second" are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0354] Furthermore, the technical solutions described in the embodiments of this application can be combined arbitrarily without conflict.

[0355] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application.

Claims

1. A channel environment prediction method, characterized in that, Applied to a management device, said management device being used to manage at least a plurality of access network devices, the method includes: Determine first information, the first information being characterized by the fact that a first model local to one or more first access network devices cannot be trained and / or that the channel environment predicted by the local first model cannot optimize network performance, the first model being used to predict the channel environment; A corresponding second model is provided for each of one or more first access network devices. The second model is used to predict the channel environment and is obtained by federated learning using multiple access network devices managed by the management device.

2. The method according to claim 1, characterized in that, The method further includes: Obtain second information about multiple access network devices, the second information including relevant information about the access network devices; Using the acquired second information, the multiple access network devices are divided into M groups, each group containing multiple access network devices, where M is an integer greater than or equal to 1; For each group, an aggregation device and multiple training devices are determined. The training devices are used to train the second model using local data and obtain training results. The training devices belong to multiple access network devices corresponding to the group. The aggregation device is used to aggregate the training results obtained by the multiple training devices to obtain the trained second model and provide the trained second model to each first access network device included in the group. Federated learning is performed using the aggregation device and multiple training devices.

3. The method according to claim 2, characterized in that, The step of dividing the multiple access network devices into M groups using the acquired second information includes: For each of the plurality of second pieces of information, feature extraction is performed on the second piece of information to obtain the feature information of the access network device; Based on the similarity of feature information of multiple access network devices, the multiple access network devices are divided into M groups.

4. The method according to claim 2, characterized in that, The second information includes information related to the local data of the access network devices. For each group, multiple training devices are identified, including: For each group, multiple training devices are determined by utilizing the computing power resource information and / or relevant information of the local data of the multiple access network devices corresponding to the group.

5. The method according to claim 2, characterized in that, The step of determining an aggregation device for each group includes: For each group, the aggregation device is determined by utilizing the computing power resource information and / or transmission latency information related to the multiple access network devices corresponding to the group, the management device, and one or more other management devices besides the management device.

6. The method according to claim 2, characterized in that, The acquisition of the second information of multiple access network devices includes: Send a fourth message to the plurality of access network devices, the fourth message being used to request the reporting of the second message; receive the second message reported by the plurality of access network devices; or, Subscribe to the second information of the multiple access network devices.

7. The method according to any one of claims 2 to 6, characterized in that, The method further includes: The fifth piece of information is determined, which indicates that one or more second access network devices are unable to optimize network performance using the channel environment predicted by the second model; The second model is retrained to obtain a retrained second model; a retrained second model is provided for each of the one or more second access network devices.

8. The method according to claim 7, characterized in that, The method further includes: Obtain the prediction performance of the plurality of access network devices in predicting the channel environment using the second model; and designate the access network devices with prediction performance lower than a first threshold as the second access network devices. or, Receive sixth information sent by one or more access network devices, the sixth information indicating that the prediction performance of the channel environment using the second model is lower than a second threshold; designate the access network device corresponding to the sixth information as the second access network device.

9. The method according to claim 7, characterized in that, The retraining of the second model includes: The second model is trained by reusing the aggregation device and multiple training devices; or, Using the acquired second information, the multiple access network devices are re-divided into N groups, where N is an integer greater than or equal to 1; for each re-divided group, an aggregation device and multiple training devices are determined; the second model is trained using the aggregation device and multiple training devices.

10. The method according to claim 9, characterized in that, In the case of overfitting of the second model, N is less than M; or in the case of underfitting of the second model, N is greater than M.

11. A channel environment prediction device, characterized in that, A management device is configured to manage at least a plurality of access network devices, including: A determining unit is configured to determine first information, the first information representing that a first model local to one or more first access network devices cannot be trained and / or the channel environment predicted by the local first model cannot optimize network performance, the first model being used to predict the channel environment; A providing unit is configured to provide a corresponding second model for each of one or more first access network devices, the second model being used to predict the channel environment, and the second model being obtained by federated learning using multiple access network devices managed by the management device.

12. A management device, characterized in that, include: The processor and the memory used to store computer programs that can run on the processor. When the processor is used to run the computer program, it performs the steps of the method according to any one of claims 1 to 10.

13. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 10.

Images