A model adjustment method, apparatus, base station and centralized node

Abstract

This application discloses a model adjustment method, apparatus, base station, and centralized node, relating to the field of communication technology, to solve the problem that traditional multi-model control schemes cannot guarantee the accuracy and speed of model switching. The method includes: predicting the second computing resource status information of the base station at a target time based on the first computing resource status information of the base station at the current time; sending the second computing resource status information to a centralized node; receiving spatial quantization raster information sent by the centralized node; dividing the coverage area space of the base station according to the spatial quantization raster information and updating the associated model input feature information; sending the updated model input feature information to the centralized node so that the centralized node determines the adjusted target model based on the model input feature information; and receiving the target model sent by the centralized node. Embodiments of this application can dynamically adjust the model size according to changes in the remaining computing resources within the base station, ensuring the accuracy and speed of model switching.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a model adjustment method, apparatus, base station and centralized node. Background Technology

[0002] In related technologies, multi-model adjustment schemes mainly adopt traditional multi-model control schemes. The principle is to treat the model as the controlled object, construct a suitable model set for the controlled object to cover the uncertainty of the controlled object, then set corresponding controllers according to each model in the model set to form a controller set, and then give a model switching strategy based on the identification error.

[0003] In wireless networks, the utilization rate of network-side resources changes dynamically. Traditional multi-model control schemes cannot allocate wireless resources in a reasonable and orderly manner. In the face of rapidly changing wireless resources, it is difficult to guarantee the accuracy and speed of model switching. Summary of the Invention

[0004] This application provides a model adjustment method, apparatus, base station, and centralized node to solve the problem that traditional multi-model control schemes cannot guarantee the accuracy and speed of model switching.

[0005] In a first aspect, embodiments of this application provide a model adjustment method, executed by a base station, the method comprising:

[0006] Based on the first computing resource status information of the base station at the current time, predict the second computing resource status information of the base station at the target time;

[0007] The second computing resource status information is sent to the centralized node so that the centralized node can determine the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information.

[0008] Receive the spatial quantization raster information sent by the centralized node;

[0009] Based on the spatial quantization grid information, the coverage area of ​​the base station is divided into spatial regions, and the associated model input feature information is updated.

[0010] The updated model input feature information is sent to the centralized node so that the centralized node can determine the adjusted target model based on the model input feature information.

[0011] Receive the target model sent by the central node.

[0012] Optionally, before predicting the second computing resource status information of the base station at the target time, the method further includes:

[0013] The computing resource status prediction module of the base station obtains wireless resource information from the wireless resource management module of the base station;

[0014] The computing resource status prediction module obtains computing task information from the computing task management module of the base station;

[0015] The computing resource status prediction module obtains the remaining computing resource information from the computing resource management module of the base station;

[0016] The first computing resource status information includes the wireless resource information, the computing task information, and the computing resource information.

[0017] Optionally, the computing resource status prediction module of the base station obtains wireless resource information from the wireless resource management module of the base station, including:

[0018] The computing resource status prediction module sends a wireless resource information request message to the wireless resource management module. The wireless resource information request message carries a first request information type, which includes at least one of uplink and downlink wireless channel quality and the number of Radio Resource Control (RRC) connections.

[0019] The computing resource status prediction module receives a wireless resource information response message returned by the wireless resource management module, wherein the wireless resource information response message carries wireless resource information corresponding to the first request information type.

[0020] Optionally, the computing resource status prediction module obtains computing task information from the computing task management module of the base station, including:

[0021] The computing resource status prediction module sends a computing task information request message to the computing task management module. The computing task information request message carries a second request information type, which includes at least one of the following: the number of active computing tasks of various types, the arrival time interval of various computing tasks, and the amount of computing resources occupied by various computing tasks.

[0022] The computing resource status prediction module receives a computing task information response message returned by the computing task management module, wherein the computing task information response message carries computing task information corresponding to the second request information type.

[0023] Optionally, the computing resource status prediction module obtains remaining computing resource information from the computing resource management module of the base station, including:

[0024] The computing resource status prediction module sends a computing resource information request message to the computing resource management module, wherein the computing resource information request message carries the requested resource type and the remaining quantity of the requested resource;

[0025] The computing resource status prediction module receives a computing resource information response message returned by the computing resource management module, wherein the computing resource information response message carries the computing resource type corresponding to the requested resource type and the remaining quantity of each type of computing resource.

[0026] Optionally, sending the second computing resource status information to the centralized node includes:

[0027] The computing resource status prediction module of the base station sends a computing resource status update message to the spatial quantization decision module of the centralized node. The computing resource status update message carries the ID of the base station, the prediction time information and the second computing resource status information.

[0028] Receiving the spatial quantized raster information sent by the centralized node includes:

[0029] The spatial quantization module of the base station receives a raster size update message sent by the spatial quantization decision module, wherein the raster size update message carries the spatial quantization raster information and the raster execution time information.

[0030] Optionally, sending the updated model input feature information to the centralized node includes:

[0031] The spatial quantization module of the base station sends an input layer update request message to the model management module of the base station. The input layer update request message carries updated model input feature information, which includes the model input feature to be updated and the updated dimension value of the model input feature.

[0032] The model management module sends a model adjustment request message to the model repository of the centralized node. The model adjustment request message carries the ID of the base station, the reason for the model update, the updated model input features, and the updated dimension value of the model input features.

[0033] Receiving the target model sent by the centralized node includes:

[0034] The model management module receives a model adjustment response message sent by the model repository of the centralized node, wherein the model adjustment response message carries the adjusted target model.

[0035] Optionally, after receiving the target model sent by the centralized node, the method further includes:

[0036] The model management module of the base station sends a model update message to the model deployment module of the base station, wherein the model update message includes the target model number and the target model parameter information;

[0037] The model deployment module deploys the target model based on the model update message.

[0038] Optionally, after the model deployment module deploys the target model according to the model update message, the method further includes:

[0039] Obtain the performance data of the base station after applying the target model;

[0040] The performance data is sent to the central node.

[0041] Secondly, embodiments of this application also provide a model adjustment method, executed by a centralized node, the method comprising:

[0042] Receive the second computational resource status information for the predicted target time sent by the base station;

[0043] Based on the second computing resource status information, determine the spatial quantization grid information within the area corresponding to the base station;

[0044] The spatial quantization raster information is sent to the base station so that the base station can divide the coverage area space of the base station according to the spatial quantization raster information and update the associated model input feature information;

[0045] Receive the model input feature information sent by the base station;

[0046] Based on the input feature information of the model, determine the adjusted target model;

[0047] The target model is sent to the base station.

[0048] Optionally, the second computational resource status information for the predicted target time sent by the receiving base station includes:

[0049] The spatial quantization decision module of the centralized node receives a computing resource status update message sent by the computing resource status prediction module of the base station, wherein the computing resource status update message carries the ID of the base station, prediction time information and the second computing resource status information.

[0050] The step of determining the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information includes:

[0051] The spatial quantization decision module determines the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information.

[0052] Sending the spatial quantization grid information to the base station includes:

[0053] The spatial quantization decision module sends a raster size update message to the spatial quantization module of the base station. The raster size update message carries the spatial quantization raster information and the raster execution time information.

[0054] Optionally, receiving the model input feature information sent by the base station includes:

[0055] The model repository of the centralized node receives a model adjustment request message sent by the model management module of the base station. The model adjustment request message carries the ID of the base station, the reason for the model update, the updated model input features, and the updated dimension value of the model input features.

[0056] Sending the target model to the base station includes:

[0057] The model repository sends a model adjustment response message to the model management module, wherein the model adjustment response message carries the adjusted target model.

[0058] Optionally, determining the adjusted target model based on the model input feature information includes:

[0059] Based on the input feature information of the model and combined with the performance data of each model stored in the model repository of the centralized node, a target model that meets the requirements is determined.

[0060] Optionally, after sending the target model to the base station, the method further includes:

[0061] Receive the performance data of the base station after applying the target model, sent by the base station;

[0062] The target model and the performance data are stored accordingly.

[0063] Optionally, receiving the performance data of the base station after applying the target model sent by the base station includes:

[0064] The centralized node's performance evaluation module receives a model performance message sent by the base station's model inference performance evaluation module. The model performance message carries the base station's ID, the target model's number, and the target model's performance data.

[0065] The corresponding storage of the target model and the performance data includes:

[0066] The effect evaluation module sends a base station model performance reward message to the model repository of the centralized node. The base station model performance reward message carries the ID of the base station, the number of the target model, and the performance data of the target model.

[0067] The model repository stores the target model and its performance data.

[0068] Optionally, the method further includes:

[0069] The performance evaluation module combines the performance data of multiple base stations after applying the target model to calculate the global inference performance data of the target model.

[0070] The performance evaluation module sends a global model performance reward message to the spatial quantization decision module of the centralized node. The global model performance reward message includes a list of IDs of the multiple base stations and the global inference performance data.

[0071] Thirdly, embodiments of this application also provide a model adjustment device, disposed in a base station, the model adjustment device comprising:

[0072] The prediction module is used to predict the second computing resource status information of the base station at a target time based on the first computing resource status information of the base station at the current time.

[0073] The first sending module is used to send the second computing resource status information to the central node, so that the central node can determine the spatial quantization grid information in the area corresponding to the base station based on the second computing resource status information.

[0074] The first receiving module is used to receive the spatial quantization raster information sent by the centralized node;

[0075] The processing module is used to divide the coverage area space of the base station according to the spatial quantization grid information and update the associated model input feature information;

[0076] The second sending module is used to send the updated model input feature information to the centralized node, so that the centralized node can determine the adjusted target model based on the model input feature information;

[0077] The second receiving module is used to receive the target model sent by the central node.

[0078] Optionally, the model adjustment device further includes:

[0079] The first acquisition module is used to acquire wireless resource information from the wireless resource management module of the base station through the computing resource status prediction module of the base station;

[0080] The second acquisition module is used to acquire computing task information from the computing task management module of the base station through the computing resource status prediction module of the base station;

[0081] The third acquisition module is used to acquire remaining computing resource information from the computing resource management module of the base station through the computing resource status prediction module of the base station;

[0082] The first computing resource status information includes the wireless resource information, the computing task information, and the computing resource information.

[0083] Optionally, the first acquisition module is used to:

[0084] The computational resource status prediction module sends a wireless resource information request message to the wireless resource management module, wherein the wireless resource information request message carries a first request information type, and the first request information type includes at least one of uplink and downlink wireless channel quality and RRC connection number;

[0085] The computing resource status prediction module receives a wireless resource information response message returned by the wireless resource management module, wherein the wireless resource information response message carries wireless resource information corresponding to the first request information type.

[0086] Optionally, the second acquisition module is used for:

[0087] The computing resource status prediction module sends a computing task information request message to the computing task management module. The computing task information request message carries a second request information type, which includes at least one of the following: the number of active computing tasks of various types, the arrival time interval of various computing tasks, and the amount of computing resources occupied by various computing tasks.

[0088] The computing resource status prediction module receives a computing task information response message returned by the computing task management module, wherein the computing task information response message carries computing task information corresponding to the second request information type.

[0089] Optionally, the third acquisition module is used for:

[0090] The computing resource status prediction module sends a computing resource information request message to the computing resource management module, wherein the computing resource information request message carries the requested resource type and the remaining quantity of the requested resource;

[0091] The computing resource status prediction module receives a computing resource information response message returned by the computing resource management module, wherein the computing resource information response message carries the computing resource type corresponding to the requested resource type and the remaining quantity of each type of computing resource.

[0092] Optionally, the first sending module is used to send a computing resource status update message to the spatial quantization decision module of the centralized node through the computing resource status prediction module of the base station, wherein the computing resource status update message carries the ID of the base station, prediction time information and the second computing resource status information.

[0093] The first receiving module is used to receive a raster size update message sent by the spatial quantization decision module through the spatial quantization module of the base station, wherein the raster size update message carries the spatial quantization raster information and the raster execution time information.

[0094] Optionally, the second transmitting module is used to:

[0095] The spatial quantization module of the base station sends an input layer update request message to the model management module of the base station. The input layer update request message carries updated model input feature information, which includes the model input feature to be updated and the updated dimension value of the model input feature.

[0096] The model management module sends a model adjustment request message to the model repository of the centralized node. The model adjustment request message carries the ID of the base station, the reason for the model update, the updated model input features, and the updated dimension value of the model input features.

[0097] The second receiving module is used to receive a model adjustment response message sent by the model repository of the centralized node through the model management module, wherein the model adjustment response message carries the adjusted target model.

[0098] Optionally, the model adjustment device further includes:

[0099] The fifth sending module is used to send a model update message to the model deployment module of the base station through the model management module of the base station, wherein the model update message includes the target model number and the parameter information of the target model;

[0100] The deployment module is used to deploy the target model according to the model update message through the model deployment module.

[0101] Optionally, the model adjustment device further includes:

[0102] The fourth acquisition module is used to acquire performance data of the base station after applying the target model;

[0103] The sixth sending module is used to send the performance data to the central node.

[0104] Fourthly, embodiments of this application also provide a model adjustment device, disposed at a centralized node, the model adjustment device comprising:

[0105] The third receiving module is used to receive the second computing resource status information for the predicted target time sent by the base station;

[0106] The first determining module is used to determine the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information;

[0107] The third sending module is used to send the spatial quantization grid information to the base station so that the base station can divide the coverage area space of the base station according to the spatial quantization grid information and update the associated model input feature information.

[0108] The fourth receiving module is used to receive the model input feature information sent by the base station;

[0109] The second determining module is used to determine the adjusted target model based on the input feature information of the model.

[0110] The fourth sending module is used to send the target model to the base station.

[0111] Optionally, the third receiving module is used to receive a computing resource status update message sent by the computing resource status prediction module of the base station through the spatial quantization decision module of the centralized node, wherein the computing resource status update message carries the ID of the base station, prediction time information and the second computing resource status information.

[0112] The first determining module is used to determine the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information through the spatial quantization decision module;

[0113] The third sending module is used to send a grid size update message to the spatial quantization module of the base station through the spatial quantization decision module. The grid size update message carries the spatial quantization grid information and the grid execution time information.

[0114] Optionally, the fourth receiving module is used to receive a model adjustment request message sent by the model management module of the base station through the model warehouse of the centralized node, wherein the model adjustment request message carries the ID of the base station, the reason for model update, the updated model input features, and the updated dimension value of the model input features;

[0115] The fourth sending module is used to send a model adjustment response message to the model management module through the model repository, wherein the model adjustment response message carries the adjusted target model.

[0116] Optionally, the second determining module is used to determine the target model that meets the requirements based on the model input feature information and the performance data of each model stored in the model warehouse of the centralized node.

[0117] Optionally, the model adjustment device further includes:

[0118] The fifth receiving module is used to receive the performance data of the base station after applying the target model, which is sent by the base station;

[0119] The storage module is used to store the target model and the performance data.

[0120] Optionally, the fifth receiving module is used to receive a model performance message sent by the model inference performance evaluation module of the base station through the effect evaluation module of the centralized node, wherein the model performance message carries the ID of the base station, the number of the target model, and the performance data of the target model;

[0121] The storage module is used for:

[0122] The effect evaluation module sends a base station model performance reward message to the model repository of the centralized node. The base station model performance reward message carries the ID of the base station, the number of the target model, and the performance data of the target model.

[0123] The model repository stores the target model and its performance data.

[0124] Optionally, the model adjustment device further includes:

[0125] The calculation module is used to calculate the global inference performance data of the target model by combining the performance data of the target model after applying it to multiple base stations through the effect evaluation module.

[0126] The seventh sending module is used to send a global model performance reward message to the spatial quantization decision module of the centralized node through the effect evaluation module. The global model performance reward message includes a list of IDs of the multiple base stations and the global inference performance data.

[0127] Fifthly, embodiments of this application also provide a communication device, including: a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the model adjustment method as described in the first aspect; or to implement the steps in the model adjustment method as described in the second aspect.

[0128] In a sixth aspect, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the model adjustment method as described in the first aspect; or implements the steps in the model adjustment method as described in the second aspect.

[0129] In this embodiment, based on the base station's first computing resource status information at the current time, a second computing resource status information of the base station at a target time is predicted; the second computing resource status information is sent to a centralized node, enabling the centralized node to determine the spatial quantization raster information within the area corresponding to the base station based on the second computing resource status information; the spatial quantization raster information sent by the centralized node is received; the coverage area space of the base station is divided according to the spatial quantization raster information, and the associated model input feature information is updated; the updated model input feature information is sent to the centralized node, enabling the centralized node to determine the adjusted target model based on the model input feature information; and the target model is received from the centralized node. Thus, by predicting the base station's computing resource status information at the target time, the centralized node adjusts the spatial quantization raster information, the base station updates the associated model input feature information, and the centralized node selects a model of corresponding size based on the new model input feature information. This achieves dynamic adjustment of the model size based on changes in the remaining computing resources within the base station, ensuring the accuracy and speed of model switching. Attached Figure Description

[0130] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0131] Figure 1 This is one of the flowcharts of the model adjustment method provided in the embodiments of this application;

[0132] Figure 2 This is the second flowchart of the model adjustment method provided in the embodiments of this application;

[0133] Figure 3 This is an interactive flowchart of each functional module in the model adjustment method provided in the embodiments of this application;

[0134] Figure 4 This is a schematic diagram of an embodiment of the joint management model of three base stations provided in this application;

[0135] Figure 5 This is one of the structural diagrams of the model adjustment device provided in the embodiments of this application;

[0136] Figure 6 This is the second structural diagram of the model adjustment device provided in the embodiments of this application;

[0137] Figure 7 This is a structural diagram of the communication device provided in the embodiments of this application. Detailed Implementation

[0138] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0139] To make the application and the technical problems solved by the embodiments of this application clearer, the technical background and related concepts involved in the embodiments of this application will be briefly introduced below:

[0140] Current solutions for multi-model tuning mainly include traditional multi-model control and model tuning based on artificial neural networks.

[0141] Traditional multi-model control (MMR) schemes treat the models as the controlled object, creating multiple models to construct a suitable model set that covers the uncertainties of the controlled object. Then, corresponding controllers are set for each model in the model set, forming a controller set. Based on the identification error, a switching strategy is provided. When the parameters of the controlled object's models change abruptly, the system can detect this change and, based on an index function, select the model closest to the controlled system model at that moment and switch to that model. Simultaneously, the control input switches to the controller corresponding to that model. This approach improves the system's transient response and reduces the risk of uncontrollable differences between models.

[0142] Model adjustments based on artificial neural networks mostly employ deep learning methods to establish a model repository. For different scenario characteristics, feature vectors are trained on the current data to extract the optimal selection method. For specific task requirements and business scenarios, the best model can be selected from the model repository as the output result.

[0143] In wireless networks, intelligent models are typically used as "plug-ins," meaning that after offline training and optimization, they are deployed to base stations for inference. Because current wireless networks lack a model repository element, adjustments cannot be made when the intelligent model structure becomes unsuitable. Model parameters can only be updated based on limited real-time data collected from within the base station.

[0144] Resource utilization within base stations in wireless networks is dynamic. Changes in the quality of the wireless channel affect the quality of the received signal, thus influencing the complexity of data transmission and reception processes, and consequently, the resource consumption of the base station. Simultaneously, changes in the service load within the wireless network directly impact the resource consumption of the base station. In future networks, each network element will possess inherent intelligence. Each base station device in a wireless network will deploy multiple intelligent models, such as deep neural networks, to infer the optimal decisions required by the base station, such as beam weights. At any given time, the more available computing resources within the base station, the smaller the size of the intelligent model used for inference, generally resulting in lower inference accuracy and shorter inference latency; conversely, the fewer available computing resources within the base station, the larger the size of the intelligent model used for inference, generally resulting in better inference accuracy but longer inference latency. How to simultaneously ensure both inference accuracy and latency performance in an environment where the available resources of the base station change dynamically has become a problem that the industry needs to research and solve.

[0145] Traditional multi-model control schemes cannot achieve reasonable and orderly allocation of wireless resources. In rapidly changing wireless resources, the accuracy and speed of model switching cannot be guaranteed. Furthermore, multi-model control schemes require constant control calculations and monitoring of each model, which imposes significant computational burden. Adjusting artificial neural network models using deep learning methods consumes substantial storage resources and struggles to fully utilize computing power. More accurate neural network model outputs require greater algorithmic complexity and longer transmission latency, making them impractical for scenarios involving multi-cell base station collaboration.

[0146] To address the aforementioned issues, this application proposes a model adjustment method based on scalable spatial quantization. It is applicable to intelligent application scenarios where the input feature dimension of the intelligent model is related to the spatial quantization granularity (hereinafter referred to as "grid") of the wireless network coverage area, such as intelligent models whose input features include user location distribution characteristics or regional signal coverage strength. This scheme dynamically adjusts the scale of the intelligent model based on changes in the remaining computing resources within the base station. The method tiers the remaining allocable resources. When all base stations in the area have sufficient remaining computing resources, the input layer feature data dimension is increased, adjusting the intelligent model to a larger scale; conversely, when computing resources are limited, the input layer feature data dimension is reduced, adjusting the intelligent model to a smaller scale. Since the grid size of all base station coverage areas must be consistent, this application's scheme first determines the uniform grid size for the area, and then selects a model of the corresponding scale based on the new input layer dimension, achieving dynamic model selection and decision processing. Furthermore, this scheme achieves a balance between computational complexity and decision algorithm accuracy, and can effectively and fully utilize computing and network resources, improving resource utilization.

[0147] See Figure 1 , Figure 1 This is a flowchart of the model adjustment method provided in the embodiments of this application, which is executed by the base station, such as... Figure 1 As shown, the method includes the following steps:

[0148] Step 101: Based on the first computing resource status information of the base station at the current time, predict the second computing resource status information of the base station at the target time.

[0149] The aforementioned computing resource status information may include one or more of the following: wireless resource information, computing task information, and information on remaining allocable computing resources.

[0150] In this embodiment, the base station can obtain its current computing resource status information and predict the computing resource status at the next moment based on the current computing resource status information. The target time for prediction can be preset according to actual needs. Specifically, the computing resource status at the next moment can be estimated based on the total resources of the base station and the current computing resource status.

[0151] Step 102: Send the second computing resource status information to the centralized node so that the centralized node can determine the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information.

[0152] In this embodiment of the application, the centralized node may refer to a network-side device, such as a network management device.

[0153] In this step, the base station can send the predicted computing resource status information at the target time, i.e., the second computing resource status information, to the central node.

[0154] The centralized node can determine the spatial quantization grid information, i.e., grid size information, within the area corresponding to the base station based on the second computing resource status information. The area corresponding to the base station can refer to the coverage area of ​​the base station or the area jointly covered by the base station and one or more other base stations.

[0155] Specifically, the centralized node can receive computing resource status information reported by one or more base stations, and make a decision on the spatial quantization granularity size (i.e., grid size decision) within the area corresponding to one or more base stations through relevant algorithms. For example, it can decide whether to use a fine-grained or coarse-grained grid size based on whether the computing resource status predicted by the base station at the target time is greater than a computing resource status threshold, so as to adjust the intelligent model to a larger or smaller model.

[0156] A grid is a granular representation of the spatial area covered by a wireless network, used for spatial quantization. For example, a grid of 1m x 1m x 1m represents the three-dimensional spatial area covered by the wireless network. Each grid's center point has location information, including longitude, latitude, and altitude. After numbering each grid, global measurement and statistics of the wireless network's characteristic variables across the entire spatial area can be achieved by measuring and statistically analyzing these variables within each grid. For instance, user location distribution can be represented by the number of users in each grid; similarly, the downlink signal coverage of the wireless network can be represented by the statistical mean of the downlink reference signal receiving power (RSRP) in each grid.

[0157] Step 103: Receive the spatial quantization raster information sent by the centralized node.

[0158] After determining the spatial quantization raster information of the corresponding area, the centralized node can send the spatial quantization raster information to all base stations in the area, so that the base stations can receive the spatial quantization raster information sent by the centralized node.

[0159] Step 104: Based on the spatial quantization grid information, divide the coverage area space of the base station and update the associated model input feature information.

[0160] In this step, the base station can redivide the coverage area space of the base station according to the grid size indicated in the spatial quantization grid information, and update the model input feature information associated with the spatial quantization grid information, such as updating the dimensions of relevant wireless network feature data, such as user location distribution features, regional signal coverage strength, etc.

[0161] Step 105: Send the updated model input feature information to the centralized node so that the centralized node can determine the adjusted target model based on the model input feature information.

[0162] After updating the model input feature information, the base station can send the updated model input feature information to the centralized node. The centralized node can then adjust the model specifications based on the model input feature information to determine the target model that conforms to the model input feature information.

[0163] Optionally, the centralized node can determine the target model that meets the requirements based on the model input feature information and the performance data of each model stored in the model repository of the centralized node.

[0164] The centralized node's model repository can store a large number of intelligent models suitable for wireless network intelligent applications, as well as performance information, i.e., performance data, after the models are deployed in a real network.

[0165] Specifically, the centralized node can determine the base station's requirements for the new model based on the model input feature information from the base station side. It can also select the target model needed by the base station and send it to the base station by combining the performance data of various models stored within it. For example, it can select the model with better performance from multiple models of corresponding sizes that match the model input features (dimensions) as the target model. This ensures that a target model that matches the base station's computing resource status and has good performance can be selected quickly and accurately.

[0166] Step 106: Receive the target model sent by the centralized node.

[0167] The base station can receive the target model sent by the centralized node, thereby obtaining a new model, and can deploy the new model to the corresponding hardware device to use the target model for inference calculation.

[0168] Optionally, before step 101, the method further includes:

[0169] The computing resource status prediction module of the base station obtains wireless resource information from the wireless resource management module of the base station;

[0170] The computing resource status prediction module obtains computing task information from the computing task management module of the base station;

[0171] The computing resource status prediction module obtains the remaining computing resource information from the computing resource management module of the base station;

[0172] The first computing resource status information includes the wireless resource information, the computing task information, and the computing resource information.

[0173] In one implementation method, such as Figure 3 As shown, the base station side is equipped with a wireless resource management module, a computing task management module, and a computing resource management module, and also has a newly added computing resource status prediction module.

[0174] The base station can obtain wireless resource information, computing task information, and remaining computing resource information from the wireless resource management module, the computing task management module, and the computing resource management module, respectively, through the computing resource status prediction module.

[0175] The wireless resource information may include, for example, uplink and downlink wireless channel quality, Radio Resource Control (RRC) connection count, and other information that affects the base station's computing resource consumption.

[0176] The computing task information may include, for example, the number of active computing tasks of various types, the arrival time interval of various computing tasks, the statistical mean and variance of the amount of computing resources occupied by various computing tasks, and may also include other information that affects the base station's computing resource consumption.

[0177] The remaining computing resource information may include, for example, the type of remaining computing resource, the statistical mean and variance of the remaining allocable amount of each type of computing resource.

[0178] This ensures that the base station can obtain relatively comprehensive computing resource status information at the current moment from different resource management modules.

[0179] Furthermore, the computing resource status prediction module of the base station obtains wireless resource information from the wireless resource management module of the base station, including:

[0180] The computing resource status prediction module sends a wireless resource information request message to the wireless resource management module. The wireless resource information request message carries a first request information type, which includes at least one of uplink and downlink wireless channel quality and RRC connection number.

[0181] The computing resource status prediction module receives a wireless resource information response message returned by the wireless resource management module, wherein the wireless resource information response message carries wireless resource information corresponding to the first request information type.

[0182] In a more specific implementation, such as Figure 3 As shown, the computing resource status prediction module can obtain wireless resource-related information from the wireless resource management module through Radio Resource Info Request / Response messages. The Request message may include: an indication of the type of information to be obtained (information type options include uplink / downlink radio channel quality, RRC connection count, and may also include other information affecting base station computing resource consumption), and the information level (options include base station level and cell level). If the information level is cell level, it may also include one or more cell numbers. The Response message may include: uplink / downlink radio channel quality statistics at the base station or cell level (such as the statistical values ​​of all users' uplink / downlink channel quality indicators (CQI) within a certain time window), and the number of RRC connections. If the information level is cell level, it may also include one or more corresponding cell numbers.

[0183] In this way, the computing resource status prediction module and the wireless resource management module can obtain the specified or required wireless resource information through a specific message interaction process.

[0184] Furthermore, the computing resource status prediction module obtains computing task information from the computing task management module of the base station, including:

[0185] The computing resource status prediction module sends a computing task information request message to the computing task management module. The computing task information request message carries a second request information type, which includes at least one of the following: the number of active computing tasks of various types, the arrival time interval of various computing tasks, and the amount of computing resources occupied by various computing tasks.

[0186] The computing resource status prediction module receives a computing task information response message returned by the computing task management module, wherein the computing task information response message carries computing task information corresponding to the second request information type.

[0187] In a more specific implementation, such as Figure 3As shown, the computing resource status prediction module can obtain computing task-related information from the computing task management module through a Compute Task Info Request / Response message. The Request message may include: an indication of the type of information to be obtained (information type options include the number of active computing tasks of each type, the arrival time interval of each type of computing task, the statistical mean and variance of the computing resources occupied by each type of computing task, and may also include other information affecting the base station's computing resource consumption), and may also include the information level (options include base station level and cell level). If the information level is cell level, it may also include one or more cell numbers. The Response message may include: the number of active computing tasks of each type, the arrival time interval of each type of computing task, and the statistical mean and variance of the computing resources occupied by each type of computing task, as indicated in the Request message, at the base station level or cell level.

[0188] The arrival time interval for various computing tasks refers to the time interval between different types of computing tasks arriving at the computing task management module.

[0189] There are many ways to classify computing tasks. For example, they can be classified according to the task objective (such as training or inference) or according to the amount of computing resources they consume (such as small-granularity tasks and large-granularity tasks).

[0190] In this way, the computing resource status prediction module and the computing task management module can obtain relevant information about specified or required computing tasks through a specific message interaction process.

[0191] Furthermore, the computing resource status prediction module obtains remaining computing resource information from the computing resource management module of the base station, including:

[0192] The computing resource status prediction module sends a computing resource information request message to the computing resource management module, wherein the computing resource information request message carries the requested resource type and the remaining quantity of the requested resource;

[0193] The computing resource status prediction module receives a computing resource information response message returned by the computing resource management module, wherein the computing resource information response message carries the computing resource type corresponding to the requested resource type and the remaining quantity of each type of computing resource.

[0194] In a more specific implementation, such as Figure 3As shown, the computing resource status prediction module can obtain information about the remaining allocable computing resources within the base station from the computing resource management module through a Compute Resource Info Request / Response message. The content of this Request message may include: an indication of the required statistical quantity (options could be the statistical mean or variance of the remaining allocable quantities of various computing resources), an indication of the required statistical unit (options could be utilization rate, floating-point operations per second (FLOPS), number of computing units, etc.), the required type of computing resource (options could be physical processor type such as Central Processing Unit (CPU), Graphics Processing Unit (GPU), Tensor Processing Unit (TPU), etc., or virtual computing resource type such as various computing units classified according to computing principles, etc.), and may also include the information level (options include base station level and cell level). If the information level is cell level, it may also include one or more cell numbers. The Response message, or Computational Resource Information Response message, may include statistics on the types and corresponding quantities of remaining allocable computing resources at the base station or cell level, as indicated in the Request message.

[0195] In this way, the computing resource status prediction module and the computing resource management module can obtain relevant information about the specified or required remaining allocable computing resources through a specific message interaction process.

[0196] Optionally, step 102 includes:

[0197] The computing resource status prediction module of the base station sends a computing resource status update message to the spatial quantization decision module of the centralized node. The computing resource status update message carries the ID of the base station, the prediction time information and the second computing resource status information.

[0198] Step 103 includes:

[0199] The spatial quantization module of the base station receives a raster size update message sent by the spatial quantization decision module, wherein the raster size update message carries the spatial quantization raster information and the raster execution time information.

[0200] In one implementation method, such as Figure 3As shown, a spatial quantization decision module can be added to the centralized node side, and a spatial quantization module can also be added to the base station side.

[0201] After collecting historical time-series data of the computing resource status information through the aforementioned interactive process, the computing resource status prediction module can predict the computing resource status at the next moment, such as... Figure 3 As shown, the predicted computing resource state for the next moment can be sent to the spatial quantization decision module on the central node side via a Compute Resource State Update message.

[0202] The computing resource status update message may include: base station ID, predicted time information (which can be absolute time or the time difference from the current time), and computing resource status information for the next time moment (which is the second computing resource status information). This computing resource status information may include the type of computing resources within the base station, the value of the remaining allocable amount of each type of computing resource (units may be utilization rate, FLOPS, or number of computing units) or its probability statistical parameters (which may be the statistical mean or statistical variance of the aforementioned quantities), or other values ​​calculated based on the computing resource status information obtained from the interaction process between the aforementioned computing resource status prediction module and the radio resource management module, computing task management module, and computing resource management module. If the above message content is reported at the cell level, it also includes the corresponding cell ID; if the computing resource status prediction module predicts the computing resource status for multiple time moments, the computing resource status update message content may include a sequence of the above information content corresponding to multiple time moments.

[0203] The spatial quantization decision module can determine the spatial quantization grid information within the area corresponding to the base station based on the information carried in the computing resource status update message, which is mainly the computing resource status information predicted by the computing resource status prediction module.

[0204] Furthermore, the spatial quantization decision module can inform all base stations within the area of ​​its decision on the spatial quantization granularity via grid size update messages. Specifically, for example... Figure 3 As shown, the spatial quantization decision module can send the grid size update message to the spatial quantization module of the base station.

[0205] The grid size update message may include: a spatially quantized grid size value (for a two-dimensional grid, this includes the grid's longitude and latitude lengths; for a three-dimensional grid, it also includes the height length) and a suggested time for grid execution (corresponding to the prediction time). If the Compute Resource StateUpdate message contains information for multiple times, then the Grid Size Update message will contain suggested grid size values ​​for multiple times.

[0206] Therefore, the spatial quantization module of the base station can receive the grid size update message sent by the spatial quantization decision module, and update the saved latest grid size value according to the grid size contained therein and the corresponding execution time. Based on this value, the coverage area of ​​the base station is re-divided, and the dimensions of relevant wireless network characteristic data, such as user location distribution characteristics and regional signal coverage strength, are updated.

[0207] In this way, the computing resource status prediction module of the base station, the spatial quantization decision module of the centralized node, and the spatial quantization module of the base station can transmit the predicted computing resource status information and spatial quantization grid information in an orderly manner through a specific message interaction process.

[0208] Optionally, step 105 includes:

[0209] The spatial quantization module of the base station sends an input layer update request message to the model management module of the base station. The input layer update request message carries updated model input feature information, which includes the model input feature to be updated and the updated dimension value of the model input feature.

[0210] The model management module sends a model adjustment request message to the model repository of the centralized node. The model adjustment request message carries the ID of the base station, the reason for the model update, the updated model input features, and the updated dimension value of the model input features.

[0211] Step 106 includes:

[0212] The model management module receives a model adjustment response message sent by the model repository of the centralized node, wherein the model adjustment response message carries the adjusted target model.

[0213] In one implementation method, such as Figure 3 As shown, the base station side can be equipped with a spatial quantization module and a model management module, and the centralized node side can also be equipped with a model repository.

[0214] The spatial quantization module of the base station can receive spatial quantization grid information sent by the centralized node, such as... Figure 3 As shown, the spatial quantization module can receive a grid size update message sent by the spatial quantization decision module of the centralized node, and re-divide the coverage area space of the base station according to the grid size indicated therein, and update the dimensions of the relevant wireless network feature data.

[0215] And as Figure 3 As shown, the spatial quantization module can inform the model management module on the base station side of the above-mentioned wireless network feature data with updated dimensions, that is, the updated model input feature information, through the Input Layer Update Request message.

[0216] The input layer update request message may include: one or more wireless network features whose dimensions need to be updated, the new dimension value of the wireless network features, the old dimension value of the wireless network features, or one or more examples of the wireless network features calculated according to the new dimension value, such as the user location distribution vector calculated according to the new dimension value of the wireless network features.

[0217] The model management module can receive the Input Layer Update Request message and search for all relevant models whose structures need updating according to the message content. It can also interact with the model repository on the centralized node side through Model Change Request / Response messages to obtain new target models. The Request message (model change request message) may include: base station ID, reason for model update (e.g., change in input layer dimension), updated model input features (mainly wireless network feature data, such as user location distribution, downlink signal coverage of a certain beam), input layer dimension requirements for the new model, and may also include relevant information about the old model (e.g., old model number, old model parameter information, or old model structure information, such as the number of old model layers, layer type of each layer, layer size parameters of each layer, etc.), and may also include structural requirements for the new model, such as the number of new model layers, layer type of each layer, layer size parameters of each layer, etc. The Response message (model change response message) may include: one or more new models that meet the requirements of the Request message, and may also include model number information.

[0218] In this way, the spatial quantization module, model management module of the base station, and model repository of the centralized node can obtain new models with matching and updated model input feature information through specific message interaction processes.

[0219] Optionally, after step 106, the method further includes:

[0220] The model management module of the base station sends a model update message to the model deployment module of the base station, wherein the model update message includes the target model number and the target model parameter information;

[0221] The model deployment module deploys the target model based on the model update message.

[0222] In one implementation method, such as Figure 3 As shown, a model management module can be added to the base station side, and a model deployment module is also set up on the base station side.

[0223] like Figure 3 As shown, after the model management module of the base station obtains the information of the target model from the model repository of the centralized node, the model management module of the base station can notify the model deployment module of the base station to deploy the new model, i.e. the target model, to the corresponding hardware device through a model update message. The model update message may include the target model number and the target model parameter information.

[0224] In this way, the model management module of the base station can complete the deployment of the adjusted target model by interacting with the model deployment module of the base station.

[0225] Furthermore, after the model deployment module deploys the target model according to the model update message, the method further includes:

[0226] Obtain the performance data of the base station after applying the target model;

[0227] The performance data is sent to the central node.

[0228] In one embodiment, the performance data of the base station after applying the target model can also be collected and sent to the central node, so that the central node can store the performance data of the target model for model selection reference, or optimize the spatial quantization granularity decision based on this feedback.

[0229] Specifically, such as Figure 3 As shown, the base station side can also be equipped with a model inference performance evaluation module and a model deployment module. The base station side also has an effect evaluation module, which is used to comprehensively evaluate the model inference performance.

[0230] The model inference performance evaluation module can obtain performance data after applying the target model from the model deployment module of the base station through the model performance metric (Model perf.metrics) message, including model inference latency, inference accuracy, etc.

[0231] The model inference performance evaluation module can send the performance data of the target model to the effect evaluation module of the central node through the model performance message. The model performance message may include: base station ID, model number, model performance index value (such as inference latency, inference accuracy, and may also include other performance index values).

[0232] The centralized node can store the target model and the performance data in correspondence for reference when selecting or adjusting the model later. Specifically, the effect evaluation module of the centralized node can send the performance data of the target model to the model repository of the centralized node. More specifically, the effect evaluation module can send the performance data of each base station after applying the new model to the model repository through the base station model performance reward (BS Model perf.Reward) message. The content of the base station model performance reward message may include: base station ID, model number, model performance data (or a reward value calculated based on the model performance index value (there may be multiple calculation algorithms)).

[0233] The model repository can store performance data of various models after application, and can analyze this performance data to form knowledge about the relationship between the internal structure of the model and the performance of the model, and use this knowledge in subsequent model selection tasks.

[0234] Optionally, the performance evaluation module can also comprehensively evaluate the performance data of all base stations after applying the new model, and calculate the global inference performance, such as... Figure 3 As shown, it can be sent to the spatial quantization decision module of the centralized node through the Global Model performance reward (GlobalModel perf.Reward) message. The content of the global model performance reward message may include: a list of base station IDs, global inference performance data (or a global reward value calculated based on the model performance index values ​​of these base stations (there may be multiple algorithms)).

[0235] Based on this feedback, the spatial quantization decision module can adjust its internal spatial quantization granularity decision algorithm. For example, when it receives the same or similar computing resource status information again, it can optimize the size of the spatial quantization granularity to make the representation of the model's input feature information more accurate.

[0236] It should be noted that the interaction message between the effect evaluation module and the spatial quantization decision module, namely the Global model performance reward (Global model perf.Reward) message, is an internal message if the two modules are located in the same physical device, and an interface message if they are located in different physical devices.

[0237] The model adjustment method of this application embodiment predicts the second computing resource status information of the base station at a target time based on the first computing resource status information of the base station at the current time; sends the second computing resource status information to a centralized node so that the centralized node can determine the spatial quantization raster information within the area corresponding to the base station based on the second computing resource status information; receives the spatial quantization raster information sent by the centralized node; divides the coverage area space of the base station according to the spatial quantization raster information and updates the associated model input feature information; sends the updated model input feature information to the centralized node so that the centralized node can determine the adjusted target model based on the model input feature information; and receives the target model sent by the centralized node. In this way, by predicting the computing resource status information of the base station at the target time, the centralized node adjusts the spatial quantization raster information, the base station updates the associated model input feature information, and the centralized node selects a model of corresponding size based on the new model input feature information. This achieves dynamic adjustment of the model size according to changes in the remaining computing resources within the base station, ensuring the accuracy and speed of model switching.

[0238] See Figure 2 , Figure 2 This is a flowchart of another model adjustment method provided in the embodiments of this application, executed by a centralized node, such as... Figure 2 As shown, the method includes the following steps:

[0239] Step 201: Receive the second computing resource status information for the predicted target time sent by the base station.

[0240] Step 202: Determine the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information.

[0241] Step 203: Send the spatial quantization raster information to the base station so that the base station can divide the coverage area space of the base station according to the spatial quantization raster information and update the associated model input feature information.

[0242] Step 204: Receive the model input feature information sent by the base station.

[0243] Step 205: Determine the adjusted target model based on the input feature information of the model.

[0244] Step 206: Send the target model to the base station.

[0245] Optionally, step 201 includes:

[0246] The spatial quantization decision module of the centralized node receives a computing resource status update message sent by the computing resource status prediction module of the base station, wherein the computing resource status update message carries the ID of the base station, prediction time information and the second computing resource status information.

[0247] Step 202 includes:

[0248] The spatial quantization decision module determines the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information.

[0249] Step 203 includes:

[0250] The spatial quantization decision module sends a raster size update message to the spatial quantization module of the base station. The raster size update message carries the spatial quantization raster information and the raster execution time information.

[0251] Optionally, step 204 includes:

[0252] The model repository of the centralized node receives a model adjustment request message sent by the model management module of the base station. The model adjustment request message carries the ID of the base station, the reason for the model update, the updated model input features, and the updated dimension value of the model input features.

[0253] Step 206 includes:

[0254] The model repository sends a model adjustment response message to the model management module, wherein the model adjustment response message carries the adjusted target model.

[0255] Optionally, step 205 includes:

[0256] Based on the input feature information of the model and combined with the performance data of each model stored in the model repository of the centralized node, a target model that meets the requirements is determined.

[0257] Optionally, after step 206, the method further includes:

[0258] Receive the performance data of the base station after applying the target model, sent by the base station;

[0259] The target model and the performance data are stored accordingly.

[0260] Optionally, receiving the performance data of the base station after applying the target model sent by the base station includes:

[0261] The centralized node's performance evaluation module receives a model performance message sent by the base station's model inference performance evaluation module. The model performance message carries the base station's ID, the target model's number, and the target model's performance data.

[0262] The corresponding storage of the target model and the performance data includes:

[0263] The effect evaluation module sends a base station model performance reward message to the model repository of the centralized node. The base station model performance reward message carries the ID of the base station, the number of the target model, and the performance data of the target model.

[0264] The model repository stores the target model and its performance data.

[0265] Optionally, the method further includes:

[0266] The performance evaluation module combines the performance data of multiple base stations after applying the target model to calculate the global inference performance data of the target model.

[0267] The performance evaluation module sends a global model performance reward message to the spatial quantization decision module of the centralized node. The global model performance reward message includes a list of IDs of the multiple base stations and the global inference performance data.

[0268] It should be noted that the embodiments in this application are as follows: Figure 1 The implementation methods shown in the embodiments corresponding to the centralized node side, including specific and optional implementation methods, can be found in [reference needed]. Figure 1 The relevant descriptions in the method embodiments shown are omitted here to avoid repetition.

[0269] The model adjustment method of this application embodiment receives second computing resource status information for a predicted target time sent by a base station; determines spatial quantization raster information within the area corresponding to the base station based on the second computing resource status information; sends the spatial quantization raster information to the base station so that the base station divides its coverage area space according to the spatial quantization raster information and updates the associated model input feature information; receives the model input feature information sent by the base station; determines the adjusted target model based on the model input feature information; and sends the target model to the base station. In this way, by adjusting the spatial quantization raster information according to the computing resource status information predicted by the base station at the target time, the base station updates the associated model input feature information, and the centralized node selects a model of corresponding size according to the new model input feature information. This achieves dynamic adjustment of the model size based on changes in the remaining computing resources within the base station, ensuring the accuracy and speed of model switching.

[0270] The following is combined Figure 4 The relevant implementation methods in the embodiments of this application are illustrated by a specific example of a joint management model of three base stations.

[0271] Figure 4 The diagram illustrates the architecture of three base stations (A, B, and C) in a wireless network, managed according to the process described in this embodiment. In this embodiment, the spatial quantization decision module is located in a centralized node (such as a network management device), and the model repository module is located in another centralized node (an independent physical device).

[0272] Base stations A, B, and C all require the deployment of intelligent models for inference to determine massive MIMO antenna weights. The input to these intelligent models is a user location distribution vector statistically based on a spatial grid; therefore, the dimension of their input features is related to the spatial quantization granularity of the wireless network coverage area. Assuming each base station has a GPU computing resource, base stations A, B, and C first monitor the status of the wireless and computing resources in each cell, obtaining the current number of computing tasks in each cell, GPU computing resource utilization, and statistical values ​​of uplink and downlink wireless channel quality and RRC connections for all users in the cell.

[0273] Base stations A, B, and C calculate the comprehensive computing resource status value T for each cell based on the collected radio and computing resource status data. The formula for calculating T is as follows:

[0274]

[0275] Where U represents GPU computing resource utilization, R represents uplink and downlink wireless signal strength, rrc represents the number of valid RRC connections, and tasks represents the current number of computing tasks. The computing resource status value T_ for the next 10 seconds is predicted based on the historical T value sequence. 10s The T_ of each cell is updated via Compute Resource State Update messages. 10s It is sent to the spatial quantization decision module in the network management device.

[0276] The network management device receives all T_ signals sent by base stations A, B, and C. 10s After setting the value, calculate the comprehensive computing resource status T′_ of all cells in this area. 10s Then determine T′_ 10s With threshold T mid The relationship can be customized by setting different spatial quantization granularity schemes, such as: when T′_ 10s >T mid When using fine-grained (e.g., 1m*1m*1m) grids to divide the cell user distribution, when T′_ 10s =T mid When using a standard granularity (3m*3m*3m) grid to divide the cell user distribution, when T′_ 10s <T mid At this time, a coarse-grained (5m*5m*5m) grid is used to divide the cell user distribution. The network management device sends the decided grid size to base stations A, B, and C through grid size update messages.

[0277] Assuming that base stations A, B, and C previously used a grid granularity of 1m*1m*3m, resulting in user location distribution vectors of 12, 18, and 24 dimensions respectively, after receiving a grid size update message from the network management device, the grid size is changed to 3m*2m*3m, making the new grid size 6 times the original grid size. Therefore, base stations A, B, and C recalculate the user location distribution vectors, with the new vectors having dimensions of 2, 3, and 4 dimensions respectively. Therefore, base stations A, B, and C need to change the input layer dimensions of the intelligent models used to decide on massive MIMO antenna weights to 2D, 3D, and 4D, respectively. Their model management modules send Model Change Request messages to the model repository, including: the new input layer dimensions (2D for base stations A, B, and C, 3D for B, and 4D, respectively), and the old intelligent model information (total number of layers (e.g., four layers), and the type of each layer (e.g., first layer: Convolutional Neural Networks (CNN) + Long Short-Term Memory (LSTM), second layer: CNN + LSTM, third layer: CNN + LSTM, fourth layer: fully connected layer). Based on the new input layer dimension information and the old model information, the model repository selects the best Convolutional Long Short-Term Memory (ConvLSTM) model that matches the current quantization level for base stations A, B, and C, and sends a Model Change Response message to base stations A, B, and C.

[0278] The model repository obtains the inference performance of new models measured in base stations A, B, and C. For example, the model performance of base station A is: inference latency: +3%, indicating that the latency standard is exceeded; coverage performance: +2%, indicating that the coverage standard is exceeded. The model performance of base station B is: inference latency: -10%, coverage performance: -5%. The model performance of base station C is: inference latency: +1%, coverage performance: +1%. Then, the performance evaluation module in the model repository calculates the comprehensive performance value of base stations A, B, and C. For example, the calculation formula is: Globalperf.=((+3%-10%+1%),(0.5×2%+0.3×-5%+0.2×1%))=(-6%, -0.3%), where 0.5, 0.3, and 0.2 are weighting coefficients. The model repository sends the comprehensive performance value (-6%, -0.3%) to the spatial quantization decision module in the network management device through the Global Model performance reward (Global Model perf.Reward) message. This global reward value indicates that the regional integrated inference latency is 6% lower than the standard, while the overall regional coverage performance is 0.3% lower than the coverage standard. Therefore, the spatial quantization decision module may adjust its internal decision algorithm based on this feedback. For example, when it receives the same or similar computing resource status information again, it may reduce the size of the spatial quantization granularity to make the user location distribution vector more accurate in representing the user distribution.

[0279] This application provides a novel solution to the problem of dynamic model adjustment in wireless networks. It can fully utilize computing and network resources, thereby improving resource utilization. It achieves a balance between computational complexity and decision algorithm accuracy. It is applicable to multi-cell joint base station scenarios and has good practical value. It is relatively simple to implement and has a certain degree of universality and applicability for different user types and distribution scenarios.

[0280] This application also provides a model adjustment device, installed in a base station. See also... Figure 5 , Figure 5 This is a structural diagram of the model adjustment device provided in the embodiments of this application. Since the principle of the model adjustment device in solving the problem is similar to that of the model adjustment method in the embodiments of this application, the implementation of the model adjustment device can refer to the implementation of the method, and the repeated parts will not be described again.

[0281] like Figure 5 As shown, the model adjustment device 500 includes:

[0282] Prediction module 501 is used to predict the second computing resource status information of the base station at a target time based on the first computing resource status information of the base station at the current time;

[0283] The first sending module 502 is used to send the second computing resource status information to the centralized node, so that the centralized node can determine the spatial quantization grid information in the area corresponding to the base station based on the second computing resource status information.

[0284] The first receiving module 503 is used to receive the spatial quantization grid information sent by the centralized node;

[0285] The processing module 504 is used to divide the coverage area space of the base station according to the spatial quantization grid information and update the associated model input feature information.

[0286] The second sending module 505 is used to send the updated model input feature information to the centralized node, so that the centralized node can determine the adjusted target model based on the model input feature information;

[0287] The second receiving module 506 is used to receive the target model sent by the centralized node.

[0288] Optionally, the model adjustment device 500 further includes:

[0289] The first acquisition module is used to acquire wireless resource information from the wireless resource management module of the base station through the computing resource status prediction module of the base station;

[0290] The second acquisition module is used to acquire computing task information from the computing task management module of the base station through the computing resource status prediction module of the base station;

[0291] The third acquisition module is used to acquire remaining computing resource information from the computing resource management module of the base station through the computing resource status prediction module of the base station;

[0292] The first computing resource status information includes the wireless resource information, the computing task information, and the computing resource information.

[0293] Optionally, the first acquisition module is used to:

[0294] The computational resource status prediction module sends a wireless resource information request message to the wireless resource management module, wherein the wireless resource information request message carries a first request information type, and the first request information type includes at least one of uplink and downlink wireless channel quality and RRC connection number;

[0295] The computing resource status prediction module receives a wireless resource information response message returned by the wireless resource management module, wherein the wireless resource information response message carries wireless resource information corresponding to the first request information type.

[0296] Optionally, the second acquisition module is used for:

[0297] The computing resource status prediction module sends a computing task information request message to the computing task management module. The computing task information request message carries a second request information type, which includes at least one of the following: the number of active computing tasks of various types, the arrival time interval of various computing tasks, and the amount of computing resources occupied by various computing tasks.

[0298] The computing resource status prediction module receives a computing task information response message returned by the computing task management module, wherein the computing task information response message carries computing task information corresponding to the second request information type.

[0299] Optionally, the third acquisition module is used for:

[0300] The computing resource status prediction module sends a computing resource information request message to the computing resource management module, wherein the computing resource information request message carries the requested resource type and the remaining quantity of the requested resource;

[0301] The computing resource status prediction module receives a computing resource information response message returned by the computing resource management module, wherein the computing resource information response message carries the computing resource type corresponding to the requested resource type and the remaining quantity of each type of computing resource.

[0302] Optionally, the first sending module 502 is used to send a computing resource status update message to the spatial quantization decision module of the centralized node through the computing resource status prediction module of the base station, wherein the computing resource status update message carries the ID of the base station, prediction time information and the second computing resource status information.

[0303] The first receiving module 503 is used to receive a grid size update message sent by the spatial quantization decision module through the spatial quantization module of the base station, wherein the grid size update message carries the spatial quantization grid information and the grid execution time information.

[0304] Optionally, the second transmitting module 505 is used for:

[0305] The spatial quantization module of the base station sends an input layer update request message to the model management module of the base station. The input layer update request message carries updated model input feature information, which includes the model input feature to be updated and the updated dimension value of the model input feature.

[0306] The model management module sends a model adjustment request message to the model repository of the centralized node. The model adjustment request message carries the ID of the base station, the reason for the model update, the updated model input features, and the updated dimension value of the model input features.

[0307] The second receiving module 506 is used to receive a model adjustment response message sent by the model warehouse of the centralized node through the model management module, wherein the model adjustment response message carries the adjusted target model.

[0308] Optionally, the model adjustment device 500 further includes:

[0309] The fifth sending module is used to send a model update message to the model deployment module of the base station through the model management module of the base station, wherein the model update message includes the target model number and the parameter information of the target model;

[0310] The deployment module is used to deploy the target model according to the model update message through the model deployment module.

[0311] Optionally, the model adjustment device 500 further includes:

[0312] The fourth acquisition module is used to acquire performance data of the base station after applying the target model;

[0313] The sixth sending module is used to send the performance data to the central node.

[0314] The model adjustment device 500 provided in this application embodiment can execute the above method embodiment, and its implementation principle and technical effect are similar, so it will not be described again here.

[0315] The model adjustment device 500 of this application embodiment predicts the second computing resource status information of the base station at a target time based on the first computing resource status information of the base station at the current time; sends the second computing resource status information to a centralized node so that the centralized node determines the spatial quantization raster information within the area corresponding to the base station based on the second computing resource status information; receives the spatial quantization raster information sent by the centralized node; divides the coverage area space of the base station according to the spatial quantization raster information and updates the associated model input feature information; sends the updated model input feature information to the centralized node so that the centralized node determines the adjusted target model based on the model input feature information; and receives the target model sent by the centralized node. In this way, by predicting the computing resource status information of the base station at the target time, the centralized node adjusts the spatial quantization raster information, the base station updates the associated model input feature information, and the centralized node selects a model of corresponding size based on the new model input feature information. This achieves dynamic adjustment of the model size according to changes in the remaining computing resources within the base station, ensuring the accuracy and speed of model switching.

[0316] This application also provides a model adjustment device, which is set at a centralized node. See also Figure 6 , Figure 6 This is a structural diagram of another model adjustment device provided in an embodiment of this application. Since the principle of the model adjustment device in solving the problem is similar to the model adjustment method in the embodiment of this application, the implementation of this model adjustment device can refer to the implementation of the method, and the repeated parts will not be described again.

[0317] like Figure 6 As shown, the model adjustment device 600 includes:

[0318] The third receiving module 601 is used to receive the second computing resource status information of the predicted target time sent by the base station;

[0319] The first determining module 602 is used to determine the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information.

[0320] The third sending module 603 is used to send the spatial quantization grid information to the base station so that the base station can divide the coverage area space of the base station according to the spatial quantization grid information and update the associated model input feature information.

[0321] The fourth receiving module 604 is used to receive the model input feature information sent by the base station;

[0322] The second determining module 605 is used to determine the adjusted target model based on the model input feature information;

[0323] The fourth sending module 606 is used to send the target model to the base station.

[0324] Optionally, the third receiving module 601 is used to receive a computing resource status update message sent by the computing resource status prediction module of the base station through the spatial quantization decision module of the centralized node, wherein the computing resource status update message carries the ID of the base station, prediction time information and the second computing resource status information.

[0325] The first determining module 602 is used to determine the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information through the spatial quantization decision module;

[0326] The third sending module 603 is used to send a grid size update message to the spatial quantization module of the base station through the spatial quantization decision module, wherein the grid size update message carries the spatial quantization grid information and the grid execution time information.

[0327] Optionally, the fourth receiving module 604 is used to receive a model adjustment request message sent by the model management module of the base station through the model warehouse of the centralized node, wherein the model adjustment request message carries the ID of the base station, the reason for model update, the updated model input features, and the updated dimension value of the model input features;

[0328] The fourth sending module 606 is used to send a model adjustment response message to the model management module through the model repository, wherein the model adjustment response message carries the adjusted target model.

[0329] Optionally, the second determining module 605 is used to determine the target model that meets the requirements based on the model input feature information and the performance data of each model stored in the model warehouse of the centralized node.

[0330] Optionally, the model adjustment device 600 further includes:

[0331] The fifth receiving module is used to receive the performance data of the base station after applying the target model, which is sent by the base station;

[0332] The storage module is used to store the target model and the performance data accordingly.

[0333] Optionally, the fifth receiving module is used to receive a model performance message sent by the model inference performance evaluation module of the base station through the effect evaluation module of the centralized node, wherein the model performance message carries the ID of the base station, the number of the target model, and the performance data of the target model;

[0334] The storage module is used for:

[0335] The effect evaluation module sends a base station model performance reward message to the model repository of the centralized node. The base station model performance reward message carries the ID of the base station, the number of the target model, and the performance data of the target model.

[0336] The model repository stores the target model and its performance data.

[0337] Optionally, the model adjustment device 600 further includes:

[0338] The calculation module is used to calculate the global inference performance data of the target model by combining the performance data of the target model after applying it to multiple base stations through the effect evaluation module.

[0339] The seventh sending module is used to send a global model performance reward message to the spatial quantization decision module of the centralized node through the effect evaluation module. The global model performance reward message includes a list of IDs of the multiple base stations and the global inference performance data.

[0340] The model adjustment device 600 provided in this application embodiment can execute the above method embodiment, and its implementation principle and technical effect are similar, so it will not be described again here.

[0341] The model adjustment device 600 of this application embodiment receives second computing resource status information for the predicted target time sent by a base station; determines spatial quantization raster information within the area corresponding to the base station based on the second computing resource status information; sends the spatial quantization raster information to the base station so that the base station divides its coverage area space according to the spatial quantization raster information and updates the associated model input feature information; receives the model input feature information sent by the base station; determines the adjusted target model based on the model input feature information; and sends the target model to the base station. In this way, by adjusting the spatial quantization raster information according to the computing resource status information predicted by the base station at the target time, the base station updates the associated model input feature information, and the centralized node selects a model of corresponding size according to the new model input feature information. This achieves dynamic adjustment of the model size based on changes in the remaining computing resources within the base station, ensuring the accuracy and speed of model switching.

[0342] This application also provides a communication device. Since the principle by which the communication device solves the problem is similar to the model adjustment method in this application, the implementation of this communication device can be found in the implementation of the method, and repeated details will not be described again. Figure 7As shown, the communication device in this application embodiment includes:

[0343] In one embodiment, the communication device is a base station. The processor 700 is used to read the program in the memory 720 and execute the following processes:

[0344] Based on the first computing resource status information of the base station at the current time, predict the second computing resource status information of the base station at the target time;

[0345] The transceiver 710 sends the second computing resource status information to the centralized node, so that the centralized node can determine the spatial quantization grid information in the area corresponding to the base station based on the second computing resource status information.

[0346] The transceiver 710 receives the spatial quantization raster information sent by the centralized node;

[0347] Based on the spatial quantization grid information, the coverage area of ​​the base station is divided into spatial regions, and the associated model input feature information is updated.

[0348] The transceiver 710 sends the updated model input feature information to the centralized node, so that the centralized node can determine the adjusted target model based on the model input feature information.

[0349] The target model is received by the transceiver 710 from the centralized node.

[0350] Transceiver 710 is used to receive and send data under the control of processor 700.

[0351] Among them, Figure 7 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 700) and memory (memory 720). The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 710 can be multiple elements, including transmitters and transceivers, providing a unit for communicating with various other devices over a transmission medium. The processor 700 is responsible for managing the bus architecture and general processing, and the memory 720 can store data used by the processor 700 during operation.

[0352] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0353] The wireless resource information is obtained from the wireless resource management module of the base station through the computing resource status prediction module of the base station.

[0354] The computing task information is obtained from the computing task management module of the base station through the computing resource status prediction module;

[0355] The remaining computing resource information is obtained from the computing resource management module of the base station through the computing resource status prediction module;

[0356] The first computing resource status information includes the wireless resource information, the computing task information, and the computing resource information.

[0357] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0358] The computational resource status prediction module sends a wireless resource information request message to the wireless resource management module, wherein the wireless resource information request message carries a first request information type, the first request information type including at least one of uplink and downlink wireless channel quality and Radio Resource Control (RRC) connection number;

[0359] The computing resource status prediction module receives a wireless resource information response message returned by the wireless resource management module, wherein the wireless resource information response message carries wireless resource information corresponding to the first request information type.

[0360] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0361] The computing resource status prediction module sends a computing task information request message to the computing task management module. The computing task information request message carries a second request information type, which includes at least one of the following: the number of active computing tasks of various types, the arrival time interval of various computing tasks, and the amount of computing resources occupied by various computing tasks.

[0362] The computing resource status prediction module receives a computing task information response message returned by the computing task management module, wherein the computing task information response message carries computing task information corresponding to the second request information type.

[0363] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0364] The computing resource status prediction module sends a computing resource information request message to the computing resource management module, wherein the computing resource information request message carries the requested resource type and the remaining quantity of the requested resource;

[0365] The computing resource status prediction module receives a computing resource information response message returned by the computing resource management module, wherein the computing resource information response message carries the computing resource type corresponding to the requested resource type and the remaining quantity of each type of computing resource.

[0366] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0367] The computing resource status prediction module of the base station sends a computing resource status update message to the spatial quantization decision module of the centralized node. The computing resource status update message carries the ID of the base station, the prediction time information and the second computing resource status information.

[0368] The base station receives a raster size update message sent by the spatial quantization decision module through its spatial quantization module. The raster size update message carries the spatial quantization raster information and the raster execution time information.

[0369] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0370] The spatial quantization module of the base station sends an input layer update request message to the model management module of the base station. The input layer update request message carries updated model input feature information, which includes the model input feature to be updated and the updated dimension value of the model input feature.

[0371] The model management module sends a model adjustment request message to the model repository of the centralized node. The model adjustment request message carries the ID of the base station, the reason for the model update, the updated model input features, and the updated dimension value of the model input features.

[0372] The model management module receives a model adjustment response message from the model repository of the centralized node, wherein the model adjustment response message carries the adjusted target model.

[0373] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0374] The model management module of the base station sends a model update message to the model deployment module of the base station, wherein the model update message includes the target model number and the target model parameter information;

[0375] The target model is deployed by the model deployment module based on the model update message.

[0376] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0377] Obtain the performance data of the base station after applying the target model;

[0378] The performance data is sent to the central node via transceiver 710.

[0379] In another embodiment, the communication device is a central node, and the processor 700 is used to read the program in the memory 720 and execute the following processes:

[0380] The transceiver 710 receives the second computing resource status information for the predicted target time sent by the base station;

[0381] Based on the second computing resource status information, determine the spatial quantization grid information within the area corresponding to the base station;

[0382] The transceiver 710 sends the spatial quantization grid information to the base station so that the base station can divide the coverage area space of the base station according to the spatial quantization grid information and update the associated model input feature information.

[0383] The transceiver 710 receives the model input feature information sent by the base station;

[0384] Based on the input feature information of the model, determine the adjusted target model;

[0385] The target model is transmitted to the base station via transceiver 710.

[0386] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0387] The spatial quantization decision module of the centralized node receives a computing resource status update message sent by the computing resource status prediction module of the base station, wherein the computing resource status update message carries the ID of the base station, prediction time information and the second computing resource status information.

[0388] The spatial quantization decision module determines the spatial quantization grid information within the area corresponding to the base station based on the second computing resource status information.

[0389] The spatial quantization decision module sends a raster size update message to the spatial quantization module of the base station. The raster size update message carries the spatial quantization raster information and the raster execution time information.

[0390] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0391] The centralized node receives a model adjustment request message sent by the model management module of the base station through its model repository. The model adjustment request message carries the ID of the base station, the reason for the model update, the updated model input features, and the updated dimension values ​​of the model input features.

[0392] The model adjustment response message is sent to the model management module through the model repository, wherein the model adjustment response message carries the adjusted target model.

[0393] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0394] Based on the input feature information of the model and combined with the performance data of each model stored in the model repository of the centralized node, a target model that meets the requirements is determined.

[0395] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0396] The transceiver 710 receives the performance data of the base station after applying the target model, which is sent by the base station.

[0397] The target model and the performance data are stored accordingly.

[0398] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0399] The centralized node's performance evaluation module receives a model performance message sent by the base station's model inference performance evaluation module. The model performance message carries the base station's ID, the target model's number, and the target model's performance data.

[0400] The effect evaluation module sends a base station model performance reward message to the model repository of the centralized node. The base station model performance reward message carries the ID of the base station, the number of the target model, and the performance data of the target model.

[0401] The model repository stores the target model and its performance data.

[0402] Optionally, the processor 700 is also used to read the program from the memory 720 and perform the following steps:

[0403] The global inference performance data of the target model is calculated by combining the performance data of multiple base stations after applying the target model with the effect evaluation module.

[0404] The effect evaluation module sends a global model performance reward message to the spatial quantization decision module of the centralized node. The global model performance reward message includes a list of IDs of the multiple base stations and global inference performance data.

[0405] The communication device provided in this application embodiment can execute the above method embodiment, and its implementation principle and technical effect are similar, so it will not be described again here.

[0406] Furthermore, the computer-readable storage medium of this application embodiment is used to store a computer program, which can be executed by a processor to implement the above. Figure 1 or Figure 2 The steps in the method embodiment shown.

[0407] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0408] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can be physically included separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0409] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute some steps of the transmission and reception methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0410] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A model adjustment method characterized by comprising: The method is executed by a base station, and comprises: According to the first calculation resource state information of the base station at the current time, predicting the second calculation resource state information of the base station at the target time, the calculation resource state information comprising wireless resource information, calculation task information and calculation resource information; Sending the second calculation resource state information to a centralized node, so that the centralized node determines the spatial quantization grid information in the area corresponding to the base station according to the second calculation resource state information; Receiving the spatial quantization grid information sent by the centralized node; According to the spatial quantization grid information, dividing the coverage area space of the base station and updating the associated model input feature information; Sending the updated model input feature information to the centralized node, so that the centralized node determines the target model after adjustment according to the model input feature information and in combination with the performance data of each model stored in the model repository of the centralized node; Receiving the target model sent by the centralized node; The method further comprises: The spatial quantization module of the base station sends an input layer update request message to the model management module of the base station, wherein the input layer update request message carries the updated model input feature information, the model input feature information comprises a model input feature to be updated and an updated dimension value of the model input feature, and the model input feature comprises wireless network feature data; The model management module sends a model adjustment request message to the model repository of the centralized node, wherein the model adjustment request message carries the ID of the base station, a model update reason, the updated model input feature and the updated dimension value of the model input feature; The method further comprises: The model management module receives a model adjustment response message sent by the model repository of the centralized node, wherein the model adjustment response message carries the target model after adjustment.

2. The method of claim 1, wherein, Before the step of predicting the second calculation resource state information of the base station at the target time, the method further comprises: The calculation resource state prediction module of the base station obtains wireless resource information from the wireless resource management module of the base station; The calculation resource state prediction module obtains calculation task information from the calculation task management module of the base station; The calculation resource state prediction module obtains residual calculation resource information from the calculation resource management module of the base station.

3. The method of claim 2, wherein, The calculation resource state prediction module of the base station obtains wireless resource information from the wireless resource management module of the base station, comprising: The calculation resource state prediction module sends a wireless resource information request message to the wireless resource management module, wherein the wireless resource information request message carries a first request information type, and the first request information type comprises at least one of uplink and downlink wireless channel quality and RRC connection number; The computing resource state prediction module receives a wireless resource information response message returned by the wireless resource management module, wherein the wireless resource information response message carries wireless resource information corresponding to the first request information type.

4. The method of claim 2, wherein, The computing resource state prediction module obtains computing task information from a computing task management module of the base station, including: The computing resource state prediction module sends a computing task information request message to the computing task management module, wherein the computing task information request message carries a second request information type, and the second request information type includes at least one of the number of active computing tasks of each type, the arrival time interval of each type of computing task, and the number of computing resources occupied by each type of computing task; The computing resource state prediction module receives a computing task information response message returned by the computing task management module, wherein the computing task information response message carries computing task information corresponding to the second request information type.

5. The method of claim 2, wherein, The computing resource state prediction module obtains residual computing resource information from a computing resource management module of the base station, including: The computing resource state prediction module sends a computing resource information request message to the computing resource management module, wherein the computing resource information request message carries a request resource type and a residual number of requested resources; The computing resource state prediction module receives a computing resource information response message returned by the computing resource management module, wherein the computing resource information response message carries a computing resource type corresponding to the request resource type and a residual number corresponding to each type of computing resource.

6. The method of claim 1, wherein, The sending of the second computing resource state information to the centralized node includes: The computing resource state prediction module of the base station sends a computing resource state update message to the space quantization decision module of the centralized node, wherein the computing resource state update message carries the ID of the base station, prediction time information, and the second computing resource state information; The receiving of the space quantization grid information sent by the centralized node includes: The space quantization module of the base station receives a grid size update message sent by the space quantization decision module, wherein the grid size update message carries the space quantization grid information and grid execution time information.

7. The method of claim 1, wherein, After the receiving of the target model sent by the centralized node, the method further includes: The model management module of the base station sends a model update message to the model deployment module of the base station, wherein the model update message includes the number of the target model and parameter information of the target model; The model deployment module deploys the target model according to the model update message.

8. The method of claim 7, wherein, After the model deployment module deploys the target model according to the model update message, the method further includes: Obtaining performance data of the base station after applying the target model; Sending the performance data to the centralized node.

9. A model adjustment method characterized by, The method is performed by a centralized node, and the method includes: receiving second computing resource state information of a predicted target time sent by a base station, the computing resource state information including wireless resource information, computing task information, and computing resource information; determining spatial quantization grid information in a region corresponding to the base station according to the second computing resource state information; sending the spatial quantization grid information to the base station to enable the base station to divide a coverage region space of the base station according to the spatial quantization grid information and update associated model input feature information; receiving the model input feature information sent by the base station; determining an adjusted target model according to the model input feature information; sending the target model to the base station; the receiving of the model input feature information sent by the base station includes: a model warehouse of the centralized node receives a model adjustment request message sent by a model management module of the base station, wherein the model adjustment request message carries an ID of the base station, a model update reason, updated model input features, and updated dimension values of the model input features, and the model input features include wireless network feature data; the sending of the target model to the base station includes: the model warehouse sends a model adjustment response message to the model management module, wherein the model adjustment response message carries the adjusted target model; the determination of the adjusted target model according to the model input feature information includes: determining a target model meeting requirements according to the model input feature information and in combination with performance data of each model stored in the model warehouse of the centralized node.

10. The method of claim 9, wherein, the receiving of second computing resource state information of a predicted target time sent by a base station includes: a spatial quantization decision module of the centralized node receives a computing resource state update message sent by a computing resource state prediction module of the base station, wherein the computing resource state update message carries an ID of the base station, prediction time information, and the second computing resource state information; the determination of spatial quantization grid information in a region corresponding to the base station according to the second computing resource state information includes: the spatial quantization decision module determines the spatial quantization grid information in the region corresponding to the base station according to the second computing resource state information; the sending of the spatial quantization grid information to the base station includes: the spatial quantization decision module sends a grid size update message to a spatial quantization module of the base station, wherein the grid size update message carries the spatial quantization grid information and grid execution time information.

11. The method of claim 9, wherein, after the sending of the target model to the base station, the method further includes: receiving performance data of the base station after the base station applies the target model; storing the target model and the performance data correspondingly.

12. The method of claim 11, wherein, the receiving of performance data of the base station after the base station applies the target model includes: The effect evaluation module of the centralized node receives a model performance message sent by the model inference performance evaluation module of the base station, wherein the model performance message carries the ID of the base station, the number of the target model, and performance data of the target model; The corresponding storage of the target model and the performance data includes: The effect evaluation module sends a base station model performance reward message to the model warehouse of the centralized node, wherein the base station model performance reward message carries the ID of the base station, the number of the target model, and the performance data of the target model; The model warehouse correspondingly stores the target model and the performance data of the target model.

13. The method of claim 12, wherein, The method further includes: The effect evaluation module combines the performance data of the target model after the target model is applied to multiple base stations to calculate global inference performance data of the target model; The effect evaluation module sends a global model performance reward message to the spatial quantization decision module of the centralized node, wherein the global model performance reward message includes an ID list of the multiple base stations and the global inference performance data.

14. A model adjustment device characterized by comprising: The model adjustment device is arranged at the base station, and the model adjustment device includes: A prediction module configured to predict second computing resource state information of the base station at a target time according to first computing resource state information of the base station at a current time, the computing resource state information including wireless resource information, computing task information, and computing resource information; A first sending module configured to send the second computing resource state information to a centralized node, so that the centralized node determines spatial quantization grid information in an area corresponding to the base station according to the second computing resource state information; A first receiving module configured to receive the spatial quantization grid information sent by the centralized node; A processing module configured to divide a coverage area space of the base station according to the spatial quantization grid information and update associated model input feature information; A second sending module configured to send the updated model input feature information to the centralized node, so that the centralized node determines an adjusted target model according to the model input feature information and in combination with performance data of each model stored in a model warehouse of the centralized node; A second receiving module configured to receive the target model sent by the centralized node; The second sending module is configured to: Send an input layer update request message to a model management module of the base station through a spatial quantization module of the base station, wherein the input layer update request message carries the updated model input feature information, the model input feature information includes a model input feature to be updated and an update dimension value of the model input feature, and the model input feature includes wireless network feature data; Send a model adjustment request message to a model warehouse of the centralized node through the model management module, wherein the model adjustment request message carries the ID of the base station, a model update reason, the updated model input feature, and the update dimension value of the model input feature; and The second receiving module is configured to: Receive the target model sent by the centralized node through the second receiving module; and Receive the model adjustment request message sent by the model management module through the second receiving module. The second receiving module is configured to receive a model adjustment response message sent by the model repository of the centralized node through the model management module, wherein the model adjustment response message carries an adjusted target model.

15. A model adjustment device characterized by comprising: The model adjustment device is arranged at the centralized node, and the model adjustment device comprises: The third receiving module is configured to receive second computing resource state information of a predicted target time sent by the base station, wherein the computing resource state information comprises wireless resource information, computing task information and computing resource information. The first determining module is configured to determine spatial quantization grid information in a region corresponding to the base station according to the second computing resource state information. The third sending module is configured to send the spatial quantization grid information to the base station, so that the base station divides a coverage region space of the base station according to the spatial quantization grid information and updates associated model input feature information. The fourth receiving module is configured to receive the model input feature information sent by the base station. The second determining module is configured to determine an adjusted target model according to the model input feature information. The fourth sending module is configured to send the target model to the base station. The fourth receiving module is configured to receive a model adjustment request message sent by a model management module of the base station through a model repository of the centralized node, wherein the model adjustment request message carries an ID of the base station, a model update reason, updated model input features and updated dimension values of the model input features, and the model input features comprise wireless network feature data. The fourth sending module is configured to send a model adjustment response message to the model management module through the model repository, wherein the model adjustment response message carries an adjusted target model. The second determining module is configured to determine a target model meeting a requirement according to the model input feature information and in combination with performance data of each model stored in the model repository of the centralized node.

16. A communication device comprising: The transceiver, the memory, the processor and the computer program stored in the memory and executable on the processor; the processor is configured to read the program in the memory to implement the steps in the model adjustment method according to any one of claims 1 to 8; or implement the steps in the model adjustment method according to any one of claims 9 to 13.

17. A computer readable storage medium for storing a computer program, characterized in that, The computer program is executed by the processor to implement the steps in the model adjustment method according to any one of claims 1 to 8; or implement the steps in the model adjustment method according to any one of claims 9 to 13.

Images