WO2024156232A1

WO2024156232A1 - Communication method and device

Info

Publication number: WO2024156232A1
Application number: PCT/CN2023/138554
Authority: WO
Inventors: 黄谢田; 曹龙雨
Original assignee: 华为技术有限公司
Priority date: 2023-01-29
Filing date: 2023-12-13
Publication date: 2024-08-02
Also published as: CN118413455A

Abstract

The present application discloses a communication method and device. The method comprises: a first communication device receives model request information, the model request information comprising inference demand information; the first communication device determines a first model according to the model request information, the first model being a multi-model; the first communication device sends first information, the first information comprising information of the first model. Therefore, after receiving the model request information, the first communication device determines a suitable multi-model (i.e., the first model) according to the inference demand information in the model request information, and then sends the information of the first model by means of the first information; after receiving the information of the first model, on the basis of the information of the first model, an inference end carries out multi-model inference by using the first model, so as to obtain an inference result having relatively high accuracy. Therefore, the method can effectively improve the use effect of a model, thereby ensuring the performance of intelligent inference (or analysis).

Description

A communication method and device

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims the priority of the Chinese patent application filed with the China Patent Office on January 29, 2023, with application number 202310115449.1 and application name “A Communication Method and Device”, the entire contents of which are incorporated by reference in this application.

Technical Field

The present application relates to the field of communication technology, and in particular to a communication method and device.

Background technique

With the continuous improvement of network intelligence and automation, the application fields of artificial intelligence (AI) and machine learning (ML) technologies are becoming more and more extensive, such as in management, core network (CN) and radio access network (RAN). Since the basic architecture of model training and reasoning has been determined and supports the use of models in various domains, how to improve the use effect of the model to ensure the performance of intelligent analysis is an issue that needs to be further explored.

In order to improve the use effect of the model, the current solution proposes that the model training functional network element (or entity) provides multiple models with different performances to the model reasoning functional network element (or entity), and the model reasoning functional network element (or entity) selects a suitable model for reasoning. However, there are some obvious defects in this solution. For example, the model reasoning functional network element (or entity) cannot determine the training capability of the model training functional network element (or entity). If multiple models are requested from the model training functional network element (or entity), the request may fail. In addition, since the number of models trained by the model training functional network element (or entity) is uncertain, and the provided model may not be suitable for the reasoning of the model reasoning functional network element (or entity), these may make the solution infeasible or the reasoning effect unsatisfactory, thereby failing to effectively improve the intelligent analysis performance of the model.

In view of the above, when using multiple models, how to obtain a suitable model for reasoning to improve model analysis performance is one of the technical problems that needs to be solved urgently.

Summary of the invention

The present application proposes a communication method and device that can effectively improve the use effect of the model to ensure the performance of intelligent reasoning (or analysis).

In a first aspect, the present application provides a communication method, which can be executed by a first communication device, or by a component of the first communication device (such as a processor, a chip, or a chip system, etc.), and the present application does not specifically limit this. The method may specifically include the following steps: the first communication device receives model request information, and the model request information includes reasoning requirement information; the first communication device determines a first model according to the model request information, and the first model is a multi-model; the first communication device sends a first message, and the first message includes information about the first model.

In the embodiment of the present application, the first communication device is regarded as the training party of the model, and the first communication device may be, but is not limited to: a model training function network element, or a model training function entity, or a communication device including a model training function. Exemplarily, the first communication device may be a NWDAF network element including a model training function module, or a network element management system (EMS) device, or an access network device (such as a base station), etc.

In the present application scheme, the first communication device receives model request information, which includes reasoning requirement information. The first communication device determines a more suitable multi-model (i.e., the first model) based on the reasoning requirement information in the model request information, and then sends the information of the first model through the first information. It can be seen that when the reasoning end (i.e., the second communication device including the model reasoning function module) receives the information of the first model, based on the information of the first model, the reasoning and combination of multiple models using the first model can obtain a more accurate reasoning result. Therefore, this method can effectively improve the use effect of the model to ensure the performance of intelligent reasoning.

In a possible implementation, before the first communication device receives the model request information, the method further includes: the first communication device sends training capability indication information, where the training capability indication information is used to indicate that the first communication device supports multi-model training.

Through this implementation, it can be ensured that the first communication device can effectively perform multi-model training after receiving the model request information.

In a possible implementation manner, when the model request information is used to request training of multiple models, the first communication device Information, determining the first model, can include the following implementation methods:

Implementation method one: the model request information also includes a multi-model training strategy; the first communication device performs training according to the inference requirement information and the multi-model training strategy to obtain multiple sub-models of the first model.

Implementation method two: The first communication device determines a multi-model training strategy based on the reasoning requirement information; and performs training based on the reasoning requirement information and the multi-model training strategy to obtain multiple sub-models of the first model.

In an embodiment of the present application, the multi-model training strategy includes one or more of the following: data processing strategy, training algorithm, training mode, number of sub-models, and type of sub-models.

Through this implementation, the first communication device can effectively train multiple sub-models of the first model based on the model request information.

In a possible implementation, when the model request information is used to request to obtain multiple models, the first communication device determines the first model according to the model request information, including: the first communication device determines the first model from at least one preset multiple model according to the inference requirement information. Exemplarily, the inference requirement information includes one or more of the following: the type of inference, the performance requirement of inference, the speed requirement of inference, and the power consumption requirement of inference.

Through this implementation, the first communication device can directly and quickly select the first model from at least one trained multiple models based on the inference requirement information.

In a possible implementation manner, the model request information also includes multi-model indication information, and the multi-model indication information is used to indicate that the model requested to be trained or obtained is a multi-model.

The multi-model indication information may also be carried in the inference requirement information included in the model request information, or the multi-model indication information may be sent separately to the first communication device, which is not specifically limited in the embodiments of the present application.

Through this implementation, the first communication device can effectively and accurately train or provide multiple models for the second communication device.

In one possible implementation, the information of the first model includes model information of the first model and information of multiple sub-models of the first model, and the information of each sub-model includes one or more of the following: identification information of the sub-model, the level of the sub-model, the performance of the sub-model, and performance constraints; the multiple sub-models include multiple first-level sub-models and one second-level sub-model, and the second-level sub-model is used to aggregate the reasoning information of the multiple first-level sub-models; or the multiple sub-models are all first-level sub-models, and the information of the first model also includes aggregation method and/or weight information.

Through this implementation, the information of multiple sub-models of the first model and the performance information of the sub-models, as well as the combination method between the reasoning information of multiple sub-models (that is, they can be combined in an aggregation manner or through second-level sub-models) can be effectively determined.

In a possible implementation manner, the method further includes: the first communication device sending reasoning performance information of the first model, where the reasoning performance information of the first model includes one or more of the following:

The performance of the first model, the size information of the first model, the power consumption of the reasoning of the first model, the reasoning speed of the first model, and the computing power of the first model.

Through this implementation, the receiving end (such as a third communication device including a model management function network element) that receives the reasoning performance information of the first model can also effectively adjust the first model based on the reasoning performance information of the first model. For example, based on the reasoning performance information of the first model and the actual reasoning requirement information, the number of sub-models of the first model can be appropriately reduced.

In a second aspect, the present application provides a communication method, which can be executed by a second communication device or by a component of the second communication device (such as a processor, a chip, or a chip system, etc.), and the present application does not specifically limit this. The method may specifically include the following steps: the second communication device receives second information, the second information includes information of a first model, the first model is determined according to reasoning requirement information, and the first model is a multi-model; the second communication device obtains reasoning information of the first model based on the information of the first model.

In the embodiment of the present application, the second communication device serves as a model reasoning party, and the second communication device may be the following but is not limited to: a model reasoning function network element, or a model reasoning function entity, or a communication device including a model reasoning function. Exemplarily, the second communication device may be a NWDAF network element including a model reasoning function module, or a network element management system (EMS) device, or an access network device (such as a base station), etc.

In the present application, the second communication device receives the information of the first model. Since the first model is a multi-model determined according to the reasoning requirement information, the second communication device uses the information of the first model and the first model to perform reasoning and combination of the multi-models to obtain a reasoning result with higher accuracy. Therefore, this method can effectively improve the use effect of the model to ensure the performance of intelligent reasoning.

In a possible implementation manner, before the second communication device receives the second information, the method further includes: the second communication device sends reasoning capability information and the reasoning requirement information; the reasoning capability information includes reasoning capability indication information, and one or more of the following: reasoning computing power and storage space; the reasoning capability indication information is used to indicate that the second communication device supports multi-model reasoning; the reasoning requirement information includes one or more of the following: the type of reasoning, the performance requirement of reasoning, the speed requirement of reasoning, and the power consumption requirement of reasoning.

Through this implementation, the second communication device sends its own reasoning capability information and reasoning requirement information, which can not only ensure that the second communication device can effectively perform multi-model reasoning in the future, but also ensure the performance of the second communication device in performing reasoning based on multiple models in the future.

In a possible implementation, the information of the first model includes model information of the first model and information of multiple sub-models of the first model, and the information of each sub-model includes one or more of the following: identification information of the sub-model, the level of the sub-model, the performance of the sub-model, and performance constraints.

Through this implementation, the second communication device can accurately obtain information about multiple sub-models of the first model, so as to effectively use these sub-models for reasoning later.

In a possible implementation, the multiple sub-models include multiple first-level sub-models and one second-level sub-model, and the second-level sub-model is used to aggregate the reasoning information of the multiple first-level sub-models; the second communication device obtains the reasoning information of the first model based on the information of the first model, including: the second communication device uses the multiple first-level sub-models to perform reasoning respectively based on the information of the multiple first-level sub-models to obtain the reasoning information of the multiple first-level sub-models; the second communication device uses the second-level sub-model to aggregate the reasoning information of the multiple first-level sub-models to obtain the reasoning information of the first model.

Through this implementation, the second communication device can use multiple first-level sub-models of the first sub-model to perform reasoning respectively, and use the second-level sub-model to effectively combine the reasoning information of the multiple first-level sub-models, so as to obtain the reasoning information of the first model.

In a possible implementation, the multiple sub-models are all first-level sub-models, and the information of the first model also includes an aggregation method and/or weight information; the second communication device obtains reasoning information of the first model based on the information of the first model, including: the second communication device uses the multiple sub-models to perform reasoning respectively based on the information of the multiple sub-models to obtain the reasoning information of the multiple sub-models; the second communication device aggregates the reasoning information of the multiple sub-models according to the aggregation method and/or weight information to obtain the reasoning information of the first model.

Through this implementation, the second communication device can also use multiple sub-models of the first sub-model to perform reasoning separately, and use specified aggregation methods and/or weight information to effectively combine the reasoning information of the multiple sub-models to obtain the reasoning information of the first model.

In a third aspect, the present application provides a communication method, which can be executed by a third communication device or by a component of the third communication device (such as a processor, a chip, or a chip system, etc.), and the present application does not specifically limit this. The method may specifically include the following steps: the third communication device receives training capability indication information of the first communication device; the training capability indication information is used to indicate that the first communication device supports multi-model training; the third communication device receives reasoning requirement information and reasoning capability information of the second communication device; the reasoning capability information includes reasoning capability indication information, and the reasoning capability indication information is used to indicate that the second communication device supports multi-model reasoning; the third communication device sends model request information to the first communication device, and the model request information includes the reasoning requirement information; the third communication device receives first information from the first communication device, the first information includes information of the first model, the first model is a multi-model, and the first model is determined according to the reasoning requirement information; the third communication device sends second information to the second communication device, and the second information includes information of the first model.

In the embodiment of the present application, the third communication device serves as a model manager, and the third communication device may be, but is not limited to: a model management function network element, or a model management function entity, or a communication device including a model management function. Exemplarily, the third communication device is a network management system (NMS) device including a model management function module.

In the present application scheme, the third communication device receives the training capability indication information of the first communication device and the reasoning requirement information and reasoning capability information of the second communication device, and determines that the first communication device supports multi-model training and the second communication device supports multi-model reasoning; then the third communication device sends a model request information carrying the reasoning requirement information to the first communication device, and the third communication device can effectively receive the first information carrying the information of the first model from the first communication device, the information of the first model is determined according to the reasoning requirement information and the first model is a multi-model; the third communication device then sends the information of the first model to the second communication device through the second information; after the second communication device receives the information of the first model, based on the information of the first model, the first model can be effectively used to perform multi-model reasoning and combination, and obtain a reasoning result with higher accuracy. Therefore, this method can effectively improve the use effect of the model to ensure the performance of intelligent reasoning.

In a possible implementation, the inference requirement information of the second communication device includes one or more of the following: the type of inference, the performance requirement of inference, the speed requirement of inference, and the power consumption requirement of inference. Through this implementation, the performance of the second communication device in subsequent inference based on multiple models can be guaranteed.

In a possible implementation manner, the reasoning capability information of the second communication device further includes one or more of the following: The computing power and storage space for reasoning. Through this implementation, it can be ensured that the second communication device effectively performs multi-model reasoning.

The multi-model indication information may also be carried in the inference requirement information included in the model request information, or the third communication device may send the multi-model indication information to the first communication device separately. This embodiment of the present application does not specifically limit this.

Through this implementation, it can be ensured that the first communication device can effectively and accurately train or provide multiple models.

In a possible implementation, the method also includes: the third communication device receives the reasoning performance information of the first model from the first communication device; the third communication device adjusts the number of sub-models in the first model according to the reasoning requirement information and the reasoning capability information of the second communication device, as well as the reasoning performance information of the first model and the information of the first model; the reasoning performance information of the first model includes one or more of the following: the performance of the first model, the size information of the first model, the power consumption of the reasoning of the first model, the reasoning speed of the first model, and the computing power of the first model.

Through this implementation, after receiving the reasoning performance information of the first model, the third communication device can adjust the number of sub-models of the first model (such as reducing the number of sub-models of the first model) based on the reasoning performance information of the first model and the reasoning requirement information and reasoning capability information of the second communication device to ensure that the reasoning performance of the sub-models of the first model actually used is better.

In the embodiment of the present application, the model information of the first model may be the identification, name, type, etc. of the first model. The identification information of the sub-model may be the storage address of the sub-model, or the unique identifier of the sub-model, etc.

Through this implementation, the manner in which the information of the multiple sub-models of the first model and the reasoning information of the sub-models are combined can be accurately determined, so that the second communication device can subsequently effectively use these sub-models to perform combined reasoning.

In a fourth aspect, an embodiment of the present application further provides a communication device, which may be the first communication device of the first aspect, or a component (e.g., a chip, or a chip system, or a circuit) in the first communication device, or a device that can be used in combination with the first communication device. In an embodiment of the present application, the first communication device may be, but is not limited to, a model training function network element, or a model training function entity, or a communication device including a model training function.

In one possible implementation, the communication device may include a module or unit corresponding to the method/operation/step/action described in the first aspect, and the module or unit may be a hardware circuit, or software, or a hardware circuit combined with software. In one possible implementation, the communication device may include a communication module (or a transceiver module) and a processing module. The processing module is used to call the communication module to perform the communication (i.e., receiving and/or sending) function.

In one possible implementation, the communication device includes a communication unit (or a transceiver unit) and a processing unit; the processing unit can be used to call the communication unit to perform communication (i.e., receiving and/or sending) functions; wherein the communication unit is used to receive model request information, and the model request information includes reasoning requirement information; the processing unit is used to determine a first model based on the model request information, and the first model is a multi-model; the communication unit is also used to send first information, and the first information includes information of the first model.

In a possible implementation, the communication unit is further used to: before receiving the model request information, send training capability indication information, where the training capability indication information is used to indicate that the first communication device supports multi-model training.

In one possible implementation, when the model request information is used to request training of multiple models, the model request information also includes a training strategy for the multiple models; when the processing unit determines the first model according to the model request information, it is specifically used to: train according to the reasoning requirement information and the training strategy for the multiple models to obtain multiple sub-models of the first model; or when the model request information is used to request training of multiple models; when the processing unit determines the first model according to the model request information, it is specifically used to: determine the training strategy for the multiple models according to the reasoning requirement information; and train according to the reasoning requirement information and the training strategy for the multiple models to obtain multiple sub-models of the first model; wherein the training strategy for the multiple models includes one or more of the following: data processing strategy, training algorithm, training mode, number of sub-models, and type of sub-models.

In a possible implementation, when the model request information is used to request acquisition of multiple models, the processing unit, when determining the first model based on the model request information, is specifically used to: determine the first model from at least one preset multiple models based on the reasoning requirement information.

In a possible implementation, the model request information also includes multi-model indication information, and the multi-model indication information is used to indicate that the model requested to be trained or obtained is a multi-model.

In a possible implementation, the reasoning requirement information includes one or more of the following: the type of reasoning, the performance requirement of reasoning, the speed requirement of reasoning, and the power consumption requirement of reasoning.

In one possible implementation, the communication unit is also used to send reasoning performance information of the first model, and the reasoning performance information of the first model includes one or more of the following: performance of the first model, size information of the first model, power consumption of reasoning of the first model, reasoning speed of the first model, and computing power of the first model.

In a fifth aspect, an embodiment of the present application further provides a communication device, which can be used for the second communication device of the second aspect, or can be a component (for example, a chip, or a chip system, or a circuit) in the second communication device, or a device that can be used in combination with the second communication device. In an embodiment of the present application, the second communication device can be, but is not limited to: a model reasoning function network element, or a model reasoning function entity, or a communication device including a model reasoning function.

In one possible implementation, the communication device may include a module or unit corresponding to the method/operation/step/action described in the second aspect, and the module or unit may be a hardware circuit, or software, or a hardware circuit combined with software. In one possible implementation, the communication device may include a processing module and a communication module (or a transceiver model). The processing module is used to call the communication module to perform the communication (i.e., receiving and/or sending) function.

In one possible implementation, the communication device includes a communication unit (or a transceiver unit) and a processing unit; wherein the communication unit receives second information, the second information includes information of a first model, the first model is determined based on reasoning requirement information, and the first model is a multi-model; the processing unit is used to obtain reasoning information of the first model based on the information of the first model.

In one possible implementation, the communication unit is also used to: send reasoning capability information and the reasoning requirement information before receiving the second information; the reasoning capability information includes reasoning capability indication information, and one or more of the following: computing power and storage space for reasoning; the reasoning capability indication information is used to indicate that the second communication device supports multi-model reasoning; the reasoning requirement information includes one or more of the following: the type of reasoning, the performance requirement of reasoning, the speed requirement of reasoning, and the power consumption requirement of reasoning.

In one possible implementation, the information of the first model includes model information of the first model and information of multiple sub-models of the first model, and the information of each sub-model includes one or more of the following: identification information of the sub-model, the level of the sub-model, the performance of the sub-model, and performance constraints.

In a possible implementation, the multiple sub-models include multiple first-level sub-models and one second-level sub-model, and the second-level sub-model is used to aggregate the reasoning information of the multiple first-level sub-models;

When obtaining the inference information of the first model based on the information of the first model, the processing unit is specifically used to: based on the information of the multiple first-level sub-models, use the multiple first-level sub-models to perform inference respectively to obtain the inference information of the multiple first-level sub-models; use the second-level sub-model to aggregate the inference information of the multiple first-level sub-models to obtain the inference information of the first model.

In one possible implementation, the multiple sub-models are all first-level sub-models, and the information of the first model also includes an aggregation method and/or weight information; when the processing unit obtains the reasoning information of the first model based on the information of the first model, it is specifically used to: based on the information of the multiple sub-models, use the multiple sub-models to perform reasoning respectively to obtain the reasoning information of the multiple sub-models; aggregate the reasoning information of the multiple sub-models according to the aggregation method and/or weight information to obtain the reasoning information of the first model.

In a sixth aspect, an embodiment of the present application further provides a communication device, which can be used for the third communication device of the third aspect, or can be a component (for example, a chip, or a chip system, or a circuit) in the third communication device, or a device that can be used in combination with the third communication device. In an embodiment of the present application, the third communication device can be, but is not limited to, a model management function network element, or a model management function entity, or a communication device including a model management function.

In a possible implementation, the communication device may include a module or unit corresponding to the method/operation/step/action described in the third aspect, and the module or unit may be a hardware circuit, or software, or a combination of a hardware circuit and software. In a possible implementation, the communication device may include a processing module and a transceiver module. The processing module is used to call the communication module (or the transceiver module) to perform the communication (ie, receiving and/or sending) function.

In one possible implementation, the communication device includes a communication unit (transceiver unit) and a processing unit; wherein the processing unit is used to call the communication unit to perform communication (i.e., receiving and/or sending) functions; the communication unit is used to receive training capability indication information of a first communication device; the training capability indication information is used to indicate that the first communication device supports multi-model training; and receive reasoning requirement information and reasoning capability information of a second communication device; the reasoning capability information includes reasoning capability indication information, and the reasoning capability indication information is used to indicate that the second communication device supports multi-model reasoning; the communication unit is also used to send model request information to the first communication device, and the model request information includes the reasoning requirement information; receive first information from the first communication device, the first information includes information of a first model, the first model is a multi-model, and the first model is determined based on the reasoning requirement information; and send second information to the second communication device, the second information includes information of the first model.

In a possible implementation, the reasoning capability information further includes one or more of the following: reasoning computing power and storage space.

In one possible implementation, the communication unit is further used to: receive reasoning performance information of the first model from the first communication device; the processing unit is further used to adjust the number of sub-models in the first model according to the reasoning requirement information and the reasoning capability information of the second communication device, as well as the reasoning information of the first model and the information of the first model; the reasoning performance information of the first model includes one or more of the following: the performance of the first model, the size information of the first model, the power consumption of the reasoning of the first model, the reasoning speed of the first model, and the computing power of the first model.

In the seventh aspect, a communication device is provided in an embodiment of the present application, and the device includes: at least one processor and an interface circuit; the interface circuit is used to provide input and/or output of programs or instructions to the at least one processor; the at least one processor is used to execute the program or instructions so that the communication device can implement the method provided by the above-mentioned first aspect or any possible implementation method thereof, or can implement the method provided by the above-mentioned second aspect or any possible implementation method thereof, or can implement the method provided by the above-mentioned third aspect or any possible implementation method thereof.

In an eighth aspect, a computer storage medium is provided in an embodiment of the present application, in which a software program is stored. When the software program is read and executed by one or more processors, the method provided by the first aspect or any possible implementation thereof can be implemented, or the method provided by the second aspect or any possible implementation thereof can be implemented, or the method provided by the third aspect or any possible implementation thereof can be implemented.

In the ninth aspect, a computer program product comprising instructions is provided in an embodiment of the present application. When the instructions are executed on a computer, the computer executes the method provided in the first aspect or any possible implementation manner thereof, or the computer executes the method provided in the second aspect or any possible implementation manner thereof, or the computer executes the method provided in the third aspect or any possible implementation manner thereof.

In the tenth aspect, a chip system is provided in an embodiment of the present application, which chip system includes a processor for supporting a device to implement the functions involved in the above-mentioned first aspect, or for supporting a device to implement the functions involved in the above-mentioned second aspect, or for supporting a device to implement the functions involved in the above-mentioned third aspect.

In a possible design, the chip system further includes a memory, and the memory is used to store necessary program instructions and data. The chip system can be composed of a chip, or can include a chip and other discrete devices.

In the eleventh aspect, a chip system is also provided in an embodiment of the present application, which includes a processor and an interface, wherein the interface is used to obtain a program or instruction, and the processor is used to call the program or instruction to implement or support the device to implement the function involved in the first aspect, or the processor is used to call the program or instruction to implement or support the device to implement the function involved in the second aspect, or the processor is used to call the program or instruction to implement or support the device to implement the function involved in the third aspect.

In a possible design, the chip system further includes a memory, the memory being used to store program instructions necessary for the terminal device. The chip system can be composed of chips or include chips and other discrete devices.

The technical effects that can be achieved by any possible implementation of the above-mentioned fourth to sixth aspects and the fourth to sixth aspects can refer to the technical effects that can be achieved by any possible implementation of the above-mentioned first to third aspects and the first to third aspects, and will not be repeated here.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of a solution flow chart for improving the use effect of the model;

FIG. 2 is an example diagram of two system logic architectures provided in an embodiment of the present application;

FIG3A is a schematic diagram of a first practical deployment architecture to which the method of an embodiment of the present application can be applied;

FIG3B is a schematic diagram of a second practical deployment architecture to which the method of the embodiment of the present application can be applied;

FIG3C is a schematic diagram of a third practical deployment architecture to which the method of the embodiment of the present application can be applied;

FIG3D is a schematic diagram of a fourth actual deployment architecture to which the method of the embodiment of the present application can be applied;

FIG3E is a schematic diagram of a fifth practical deployment architecture to which the method of the embodiment of the present application can be applied;

FIG4A is a flow chart of a communication method provided in an embodiment of the present application;

FIG4B is a schematic diagram of a flow chart of another communication method provided in an embodiment of the present application;

FIG4C is an example diagram of a multi-model training and reasoning process provided in an embodiment of the present application;

FIG5 is a schematic diagram of a flow chart of a first embodiment provided in the embodiments of the present application;

FIG6 is a schematic diagram of a flow chart of a second embodiment provided in the present application;

FIG7 is a schematic diagram of a flow chart of a third embodiment provided in the embodiments of the present application;

FIG8 is a schematic diagram of a flow chart of a fourth embodiment provided in the embodiments of the present application;

FIG9 is a schematic diagram of a flow chart of a fifth embodiment provided in the embodiments of the present application;

FIG10 is a schematic diagram of a flow chart of a sixth embodiment provided in an embodiment of the present application;

FIG11 is a schematic diagram of a flow chart of a seventh embodiment provided in the embodiments of the present application;

FIG12 is a schematic diagram of the structure of a communication device provided in an embodiment of the present application;

FIG13 is a schematic diagram of the structure of another communication device provided in an embodiment of the present application;

FIG. 14 is a schematic diagram of the device structure of a chip provided in an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings. The specific operation method in the method embodiment can also be applied to the device embodiment or the system embodiment. In the description of the present application, unless otherwise specified, the meaning of "multiple" is two or more.

With the continuous improvement of network intelligence and automation, the application fields of artificial intelligence (AI) and machine learning (ML) technologies are becoming more and more extensive, mainly including management domain, core network (CN) domain and radio access network (RAN) domain.

In the intelligentization of the management domain, the current protocol defines a management data analytics service (MDAS). MDAS producers can process and analyze data related to network, service events and status based on AI/ML technology, and provide analysis reports for network and service operations. In the intelligentization of the core network domain, the network data analysis function (NWDAF) can perform network data analysis based on ML models, obtain data analysis results, and provide them to the network, network management and applications for policy decision-making. In the intelligentization of the RAN domain, the current main research is to support the definition of the functional framework of RAN intelligence, that is, to make necessary enhancements based on the current RAN architecture and interfaces to support network intelligence.

Since the basic architecture of model training and reasoning in the current standard has been determined accordingly and supports the use of models in various fields, how to improve the use effect of the model to ensure the performance of intelligent analysis is an issue that needs further discussion.

In a scheme for improving the use effect of the model, it is proposed that the model training functional network element (or entity) provides a plurality of models with different performances to the model reasoning functional network element (or entity), and the model reasoning functional network element (or entity) selects a suitable model for reasoning. Exemplarily, as shown in FIG1 , the steps for implementation include: S101: the model reasoning functional entity sends a model request to the model training functional entity, and the model request includes the reasoning type (such as coverage problem analysis, cell traffic prediction, etc.) and performance requirements (such as precision, accuracy, etc.); S102: the model training functional entity trains multiple models according to the reasoning type and performance requirements in the model request; S103: the model The model training functional entity sends a model response to the model reasoning functional entity, where the model response includes a reasoning type and a list of multiple models, for example, <model ID 1, model ID 2, …>.

However, the above scheme has some obvious defects. For example, the model reasoning functional entity cannot determine the training capability of the model training functional entity. If multiple models are requested from the model training functional entity, the request may fail. In addition, since the number of models trained by the model training functional entity is uncertain and the provided models are not necessarily suitable for the reasoning of the model reasoning functional entity, these may lead to the scheme being infeasible or the reasoning effect being unsatisfactory, thereby failing to effectively improve the intelligent reasoning (or analysis) performance of the model.

Therefore, the present application proposes a communication method that can effectively improve the use effect of the model, thereby ensuring the performance of intelligent reasoning. The method proposed in the present application can be applied to the 5G system architecture, and can also be applied to but not limited to the long-term evolution (LTE) communication system, and various wireless communication systems that will evolve in the future.

The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

The system logical architecture applicable to the embodiments of the present application mainly includes a model management functional entity, a model training functional entity and a model reasoning functional entity. Figure 2 shows example diagrams of several system logical architectures of the embodiments of the present application. As shown in (1) in Figure 2, the model management functional entity, the model training functional entity and the model reasoning functional entity can all be independent logical entities, and there is an interface for communication between any two entities; as shown in (2) in Figure 2, the model management functional entity is optional, and the model management functional entity can be set inside the model training functional entity. The model training functional entity and the model reasoning functional entity are independent logical entities, and there is an interface for communication between the two.

In addition, the present application also provides several practical deployment architectures to which the method of the embodiment of the present application can be applied, as shown in Figures 3A-3E. Among them, Figure 3A shows an architecture that supports model reasoning in the management domain, the model management function entity can be deployed in the network management system (network management system, NMS) equipment, the model training function entity and the model reasoning function entity can be deployed in the element management system (element management system, EMS) equipment. Figure 3B shows an architecture that supports model reasoning in the RAN domain, the model management function entity can be deployed in the NMS, the model training function entity can be deployed in the EMS, and the model reasoning function entity can be deployed in the RAN domain equipment, such as a base station. FIG3C shows an architecture supporting model reasoning in the CN domain. The model training functional entity can be deployed in the NWDAF (MTLF), and the model reasoning functional entity can be deployed in the NWDAF (AnLF). The network repository function (NRF) is mainly used for the management of network functions (NFs), including NF registration/update/deregistration, NF discovery, etc. Since the NWDAF can be regarded as a NF, the NRF can manage the NWDAF. FIG3D and FIG3E respectively show an architecture supporting air interface-related model reasoning in the RAN domain. In FIG3D, the model training functional entity can be deployed in the base station, and the model reasoning functional entity can be deployed in the user equipment (UE); optionally, the model reasoning functional entity can also be deployed in the base station when both sides can support reasoning. In FIG3E, the model training functional entity and the model reasoning functional entity are deployed on both the base station and the UE side, which can be applied to the scenario of joint training of the base station and the UE in the bilateral model scenario.

The functions of the network elements/modules/devices in the above actual deployment architecture are introduced below.

Network Management System NMS: NMS can also be called a cross-domain management system, which is responsible for the operation, management and maintenance of the network.

Element Management System (EMS): EMS can also be called a domain management system or a single domain management system, which is used to manage one or more network elements of a certain category.

The above-mentioned NMS and EMS can also be collectively referred to as 3GPP management system, or operations administration and maintenance (OAM) module.

Radio Access Network (RAN): provides wireless access to user devices, allowing users to access the network.

Core network CN: mainly provides user connection, user management and service carrying, and provides an interface to the external network as a bearer network.

Network data analysis function NWDAF network element: responsible for data analysis in the core network domain, such as detecting abnormal user behavior and analyzing slice load.

Model management functional entity: responsible for model-related lifecycle management, including training strategy configuration, etc.

Model training functional entity: responsible for model training and generating an ML model after model training is completed.

Model reasoning functional entity: responsible for the reasoning of the model, using the model to obtain the reasoning output or reasoning result.

AnLF: is the analysis logic function of the NWDAF network element.

MTLF: Model training logic function for NWDAF network elements.

Base station: A device in a mobile communication system that connects the fixed part with the wireless part and is connected to mobile terminals through wireless channels in the air.

UE: User terminal equipment, a device that allows users to access the network.

The deployment architecture shown in Figures 3A-3E is not limited to including only the entities shown in the figures, but may also include other devices not shown in the figures, which will not be listed one by one in this application.

The model management functional entity, model training functional entity, and model reasoning functional entity included in the NMS, EMS, RAN, UE, and NWDAF in Figures 3A-3E can also be called model management function, model training function, model reasoning function, and can also be called model management function network element/module, model training function network element/module, and model reasoning function network element/module.

In the embodiments of the present application, the network element or function can be a network element in a hardware device, a software function running on dedicated hardware, or a virtualized function instantiated on a platform (e.g., a cloud platform). As a possible implementation method, the above network element or function can be implemented by one device, or by multiple devices together, or can be a functional module in one device, which is not specifically limited in the embodiments of the present application.

For the convenience of explanation, the embodiments of the present application are explained by taking the network element as an example, and the XX network element is directly referred to as XX, for example, the SMF network element is referred to as SMF. It should be understood that the names of all network elements in the present application are only examples, and they may be called other names in future communications, or the network elements involved in the present application may be replaced by other entities or devices with the same functions in future communications, and the present application does not limit this. A unified explanation is given here, and no further description will be given later.

It should be noted that the names of all messages and information in this application are only examples and may be other names, which are not limited in this application. It should be understood that the message or information from network element 1 to network element 2 may be a message sent directly from network element 1 to network element 2, or may be sent indirectly, for example, network element 1 first sends a message to network element 3, and network element 3 then sends a message to network element 2, and finally the message or information is sent to network element 2 through one or more network elements.

In the embodiments of the present application, the meanings of "reasoning" and "analysis" can be regarded as the same, and both are implemented based on models. For example, "performing reasoning based on multiple models to obtain reasoning results" is equivalent to "performing analysis based on multiple models to obtain analysis results", "reasoning type" is equivalent to "analysis type", and "reasoning identifier" is equivalent to "analysis identifier".

In addition, in this application, "indication" may include direct indication, indirect indication, explicit indication, and implicit indication. When describing that a certain indication information is used to indicate A, it can be understood that the indication information carries A, directly indicates A, or indirectly indicates A.

In this application, the information indicated by the indication information is referred to as the information to be indicated. In the specific implementation process, there are many ways to indicate the information to be indicated, such as but not limited to, the information to be indicated can be directly indicated, such as the information to be indicated itself or the index of the information to be indicated. The information to be indicated can also be indirectly indicated by indicating other information, wherein there is an association relationship between the other information and the information to be indicated. It is also possible to indicate only a part of the information to be indicated, while the other parts of the information to be indicated are known or agreed in advance. For example, the indication of specific information can also be achieved with the help of the arrangement order of each information agreed in advance (such as specified by the protocol), thereby reducing the indication overhead to a certain extent.

The information to be indicated can be sent as a whole, or divided into multiple sub-information and sent separately, and the sending period and/or sending time of these sub-information can be the same or different. The specific sending method is not limited in this application. Among them, the sending period and/or sending time of these sub-information can be pre-defined, for example, pre-defined according to the protocol, or configured by the transmitting device by sending configuration information to the receiving device. Among them, the configuration information can include, for example, but not limited to, one or a combination of at least two of radio resource control signaling, media access control (media access control, MAC) layer signaling and physical layer signaling. Among them, radio resource control signaling, for example, radio resource control (radio resource control, RRC) signaling; MAC layer signaling, for example, includes MAC control element (control element, CE); physical layer signaling, for example, includes downlink control information (downlink control information, DCI).

The technical solution of the present application is introduced below in conjunction with specific embodiments.

An embodiment of the present application provides a communication method, which can be applied to but not limited to the actual deployment architecture shown in Figures 3A-3E, and the method can be executed by the network element involved in the present application, or by the chip corresponding to the network element involved. The network element in the present application can be a physical entity network element or a virtual network element. The present application does not specifically limit the form of the network element involved.

FIG4A is a flow chart of a communication method proposed in an embodiment of the present application. The method can be executed by a transceiver and/or processor of a first communication device (or a second communication device or a third communication device), or by a chip corresponding to the transceiver and/or processor. Alternatively, the embodiment can also be implemented by a controller or control device connected to the first communication device (or a second communication device or a third communication device), and the controller or control device is used to manage at least one device including the first communication device (or a second communication device or a third communication device). And the present application does not make specific restrictions on the specific form of the communication device that executes this embodiment. In addition, it should be noted that the ordinal numbers such as "first" and "second" mentioned below are used to distinguish multiple objects for the purpose of description, and are not used to limit the order, timing, priority or importance of multiple objects. Please refer to FIG4A for the specific flow of the method. as follows:

S401A: The third communication device receives training capability indication information from the first communication device, where the training capability indication information is used to indicate that the first communication device supports multi-model training.

In the embodiment of the present application, the first communication device can be used as a model training end (or model determination end), and the first communication device can be, but is not limited to: a model training function network element, or a model training function entity, or a communication device including a model training function, such as a NWDAF network element including a model training function module, or a network element management system (EMS) device, or an access network device (such as a base station), etc. The third communication device can be, but is not limited to: a model management function network element, or a model management function entity, or a communication device including a model management function, such as a network management system (NMS) device including a model management function module.

In some embodiments, the third communication device may first send training capability query information to the first communication device, and after the first communication device receives the training capability query information, it may send the training capability indication information to the third communication device. In other embodiments, the first communication device may also actively report (i.e., send) the training capability indication information to the third communication device.

S402A: The third communication device receives reasoning requirement information and reasoning capability information of the second communication device, where the reasoning capability information includes reasoning capability indication information, and the reasoning capability indication information is used to indicate that the second communication device supports multi-model reasoning.

In an embodiment of the present application, the second communication device can serve as a model reasoning end (or model usage end), and the second communication device can be but is not limited to: a model reasoning function network element, or a model reasoning function entity, or a communication device including a model reasoning function, such as a NWDAF network element including a model reasoning function module, or a network element management system (EMS) device, or an access network device (such as a base station), etc.

In addition, the first communication device, the second communication device, and the third communication device can all be independent devices, or the first communication device, the second communication device, and the third communication device can be respectively located in independent devices; or the first communication device and the second communication device are located in the same device; or the first communication device and the second communication device are the same device; therefore, the embodiment of the present application does not specifically limit the specific form of the first communication device, the second communication device, and the third communication device, the device where each communication device is located, and the location.

In some embodiments, the third communication device first sends reasoning requirement query information and reasoning capability query information to the second communication device. After the second communication device receives the reasoning requirement query information and reasoning capability query information, it sends the reasoning requirement information and the reasoning capability query information to the third communication device.

The above-mentioned reasoning requirement query information and reasoning capability query information can be sent separately by the third communication device, or they can be sent simultaneously, that is, the time when the third communication device sends the reasoning requirement query information and the reasoning capability query information is not limited. In addition, the reasoning requirement query information and the reasoning capability query information can be carried in the same message and sent by the third communication device, or they can be carried in different messages and sent by the third communication device, and the embodiments of the present application do not limit this.

In other embodiments, the second communication device may also actively report reasoning requirement information and reasoning capability information to the third communication device, where the reasoning capability information includes reasoning capability indication information, and the reasoning capability indication information is used to indicate that the second communication device supports multi-model reasoning.

Similarly, the above-mentioned reasoning requirement information and reasoning capability information can be sent separately by the second communication device, or sent simultaneously by the second communication device, that is, the time when the second communication device sends the reasoning requirement information and the reasoning capability information is not limited. In addition, the reasoning requirement information and the reasoning capability information can be carried in the same message and sent by the second communication device, or they can be carried in different messages and sent by the second communication device, and the embodiments of the present application do not limit this.

In an embodiment of the present application, the reasoning requirement information of the second communication device may include but is not limited to one or more of the type of reasoning (for example, coverage problem analysis, cell traffic prediction, etc.), performance requirements of reasoning (such as precision, accuracy, etc.), speed requirements of reasoning, and power consumption requirements of reasoning.

In the embodiment of the present application, there is no specific limitation on the order of executing the above steps S401A and S402A.

S403A: The third communication device sends model request information to the first communication device, wherein the model request information includes the inference requirement information. Correspondingly, the first communication device receives the model request information.

In a possible implementation, the model request information also includes multi-model indication information, where the multi-model indication information is used to indicate that the model requested to be trained or obtained is a multi-model.

In an embodiment of the present application, the multi-model indication information may also be carried in the inference requirement information included in the model request information, or the third communication device may separately send the multi-model indication information to the first communication device. The present application does not make any specific limitations on this.

S404A: The first communication device determines a first model according to the model request information, where the first model is a multi-model.

In one embodiment, when the model request information is used to request training of multiple models, the first communication device, based on the model request information, Determining the first model may include but is not limited to the following methods:

Method 1: If the model request information also includes a multi-model training strategy, the first communication device can perform training according to the inference requirement information and the multi-model training strategy to obtain multiple sub-models of the first model.

Method 2: The first communication device first determines a multi-model training strategy based on the inference requirement information; and then performs training based on the inference requirement information and the multi-model training strategy to obtain multiple sub-models of the first model.

In an embodiment of the present application, the multi-model training strategy may include but is not limited to data processing strategy, training algorithm, training mode, number of sub-models, type of sub-models, etc.

In another embodiment, when the model request information is used to request to obtain multiple models, the first communication device determines the first model according to the model request information, which may include: the first communication device determines the first model from at least one preset multiple model according to the inference requirement information. In an embodiment of the present application, the at least one preset multiple model may be a multiple model that has been trained in advance by the first communication device.

S405A: The first communication device sends first information, where the first information includes information of the first model.

Correspondingly, the third communication device receives the first information from the first communication device, and the information of the first model includes model information of the first model and information of multiple sub-models of the first model. The information of each sub-model includes but is not limited to identification information of the sub-model, the level of the sub-model, the performance of the sub-model, and performance constraints.

Exemplarily, the identification information of the submodel may be, but is not limited to: the name (identification) of the submodel, the storage address information of the submodel, and the unique identification number of the submodel. The level of the submodel may include a first-level submodel and a second-level submodel; wherein the second-level submodel may be used to aggregate the reasoning information of multiple first-level submodels or multiple first-level submodels.

In some embodiments, the multiple sub-models of the first model include multiple first-level sub-models and one second-level sub-model, and the second-level sub-model is used to aggregate the reasoning information of the multiple first-level sub-models.

In other embodiments, the multiple sub-models of the first model are all first-level sub-models, and the information of the first model also includes aggregation method and/or weight information.

In one embodiment, the first communication device also sends the reasoning performance information of the first model to the third communication device, and correspondingly, the third communication device receives the reasoning performance information of the first model; the reasoning performance information of the first model may include but is not limited to the performance of the first model, the size information of the first model, the power consumption of the reasoning of the first model, the reasoning speed of the first model, the computing power of the first model, etc.

S406A: The third communication device sends second information to the second communication device, where the second information includes information of the first model.

Correspondingly, the second communication device receives the second information.

In one implementation, before the third communication device executes step S406A, the third communication device further executes the following steps:

The third communication device adjusts the number of sub-models in the first model according to the reasoning requirement information and the reasoning capability information of the second communication device, as well as the reasoning information of the first model and the information of the first model. By adjusting the number of sub-models of the first model, it can be ensured that the effect of actually using the first model is better.

S407A: The second communication device obtains inference information of the first model based on the information of the first model.

In one implementation, the multiple sub-models of the first model include multiple first-level sub-models and one second-level sub-model; the second communication device obtains the reasoning information of the first model based on the information of the first model, which may include:

The second communication device uses the multiple first-level sub-models to perform reasoning respectively based on the information of the multiple first-level sub-models to obtain reasoning information of the multiple first-level sub-models; the second communication device then uses the second-level sub-model to aggregate the reasoning information of the multiple first-level sub-models to obtain the reasoning information of the first model.

In another embodiment, when the multiple sub-models of the first model are all first-level sub-models, and the information of the first model also includes aggregation method and/or weight information; the second communication device obtains the reasoning information of the first model based on the information of the first model, which may include: the second communication device uses the multiple sub-models to perform reasoning respectively based on the information of the multiple sub-models to obtain the reasoning information of the multiple sub-models; the second communication device then aggregates the reasoning information of the multiple sub-models according to the aggregation method and/or weight information to obtain the reasoning information of the first model.

In summary, in the embodiment shown in FIG4A, after the first communication device receives the model request information, it can determine a more appropriate multi-model (i.e., the first model) according to the reasoning requirement information in the model request information, and then send the information of the first model through the first information; when the reasoning end (i.e., the second communication device including the model reasoning function module) receives the information of the first model, based on the information of the first model, the multi-model reasoning and combination are performed using the first model, and a reasoning result with higher accuracy can be obtained. Therefore, the use effect of the model can be effectively improved, and the performance of intelligent reasoning (or analysis) can be guaranteed.

Figure 4B is a flow chart of another communication method proposed in an embodiment of the present application. The method can be executed by a transceiver and/or processor of a first communication device (or a second communication device), or by a chip corresponding to the transceiver and/or processor. Alternatively, the embodiment can also be implemented by a controller or control device connected to the first communication device (or a second communication device), and the controller or control device is used to manage at least one device including the first communication device (or a second communication device). And the present application does not make specific restrictions on the specific form of the communication device that executes this embodiment. Please refer to Figure 4B, the specific process of the method is as follows:

S401B: The first communication device receives model request information from the second communication device, where the model request information includes reasoning requirement information.

In an embodiment of the present application, the reasoning requirement information may include, but is not limited to, the type of reasoning, the performance requirement of reasoning, the speed requirement of reasoning, and the power consumption requirement of reasoning.

In an embodiment of the present application, the first communication device may be used as a model training end (or a model determination end), and the first communication device may be, but is not limited to: a model training function network element, or a model training function entity, or a communication device including a model training function, such as a NWDAF network element including a model training function module, or a network element management system (EMS) device, or an access network device (such as a base station), etc. The second communication device may be used as a model reasoning end (or a model use end), and the second communication device may be, but is not limited to: a model reasoning function network element, or a model reasoning function entity, or a communication device including a model reasoning function, such as a NWDAF network element including a model reasoning function module, or a network element management system (EMS) device, or an access network device (such as a base station), etc.

In addition, the first communication device and the second communication device can be independent devices, or the first communication device and the second communication device can be located in independent devices respectively; or the first communication device and the second communication device are located in the same device; or the first communication device and the second communication device are the same device; therefore, the embodiment of the present application does not specifically limit the specific form of the first communication device and the second communication device, the device where each communication device is located, and the location.

In some embodiments, before the first communication device receives the model request information from the second communication device, the first communication device may first send training capability indication information to the second communication device, and the training capability indication information is used to indicate that the first communication device supports multi-model training. In other embodiments, the first communication device may also receive reasoning capability information from the second communication device, and the reasoning capability information includes but is not limited to: reasoning capability indication information, and the reasoning capability indication information is used to indicate that the second communication device supports multi-model reasoning. Optionally, the reasoning capability information may also include the computing power and storage space for reasoning.

S402B: The first communication device determines a first model according to the model request information, where the first model is a multi-model.

In the embodiment of the present application, when the first communication device executes the step S402B, specific reference may be made to the above-mentioned step S404A, which will not be repeated here.

S403B: The first communication device sends first information, the first information including information of the first model. Correspondingly, the second communication device receives the first information (ie, second information).

In an embodiment of the present application, the information of the first model includes model information of the first model and information of multiple sub-models of the first model. The information of each sub-model includes but is not limited to identification information of the sub-model, the level of the sub-model, the performance of the sub-model, and performance constraints.

Exemplarily, the identification information of the submodel may be, but is not limited to: the name (identification) of the submodel, the storage address information of the submodel, and the unique identification symbol of the submodel. The level of the submodel may include a first-level submodel and a second-level submodel; wherein the second-level submodel may be used to aggregate the reasoning information of multiple first-level submodels or multiple first-level submodels.

S404B: The second communication device obtains inference information of the first model based on the information of the first model.

In the embodiment of the present application, when the second communication device executes the step S404B, specific reference may be made to the above-mentioned step S407A, which will not be repeated here.

For example, FIG4C shows an example diagram of a multi-model training and reasoning process proposed in an embodiment of the present application. In FIG4C, the model training function will start multiple learner training models according to the reasoning requirement information of the model reasoning function. The training method may include: the model training function first splits the original data set into multiple sub-data sets according to a certain strategy, that is, data set 1, data set 2...data set n, where n is a positive integer; then, the model training function uses each sub-data set to train the sub-learner, thereby obtaining multiple sub-learner models, that is, sub-learner 1, sub-learner 2...sub-learner n in the figure. After the model reasoning function obtains the information of the multiple sub-learner models from the model training function, the reasoning data can be input into the multiple sub-learners (that is, sub-learner 1, sub-learner 2...sub-learner n) according to the information of the multiple sub-learner models to obtain corresponding reasoning outputs, that is, reasoning output 1, reasoning output 2...reasoning output n; then, the model reasoning function combines these reasoning outputs according to a preset aggregation method (such as voting method, simple average method, weighted average method, linear hybrid method) to obtain the final reasoning output.

In summary, in the embodiments shown in FIG. 4B and FIG. 4C , when the first communication device supports multi-model training, requesting the multi-model from the first communication device can avoid the failure of the multi-model request, and the first communication device can determine (or train) a suitable multi-model that meets the reasoning requirements based on the reasoning requirements of the second communication device. Therefore, in the use of the multi-model, the method can obtain Appropriate multiple models are used to implement reasoning, thereby improving the reasoning (or analysis) performance of the model.

In the following specific embodiments, a communication method proposed by the above-mentioned solution of the present application is further elaborated in detail for different application scenarios.

Embodiment 1:

In this first embodiment, the solution of the present application is applied to the deployment architecture shown in FIG. 3A above, by enhancing the model training and deployment process of the OAM (NMS/EMS) domain to support reasoning and analysis of the management domain based on the combination of multiple learner models. The first communication device and the second communication device in the solution of the present application are the same network element management system device (EMS for short) in FIG. 3A that includes a model training function module and a model reasoning function module, and the third communication device in the solution of the present application is a network management system device (NMS for short) in FIG. 3A that includes a model management function module. Referring to FIG. 5, the specific process of the first embodiment is as follows:

S501a: The model management function module in the NMS sends reasoning requirement query information to the model reasoning function module in the EMS.

Correspondingly, the model reasoning function module in the EMS receives the reasoning requirement query information, and the reasoning requirement query information is used to query the reasoning requirement of the model reasoning function module.

S501b: The model reasoning function module in the EMS sends reasoning requirement information to the model management function module in the NMS.

Correspondingly, the model management function module in the NMS receives the reasoning requirement information.

In one implementation, the above step S501a may also be omitted, that is, the model management function module in the NMS does not send the reasoning requirement query information to the model reasoning function module in the EMS, but the model reasoning function module in the EMS actively reports (i.e., sends) the reasoning requirement information to the model management function module in the NMS.

The above-mentioned reasoning requirement information may include: reasoning type requirements (such as coverage problem analysis, cell traffic prediction, etc.), reasoning performance requirements (such as reasoning precision, accuracy, etc.), reasoning speed requirements (optional), and reasoning power consumption requirements (optional). Among them, the reasoning speed requirement can also be called the reasoning delay requirement, which indicates the requirement for the time to execute reasoning, for example, the time for a single reasoning execution is less than 1s; the reasoning power consumption requirement indicates the requirement for the power consumption of executing reasoning, for example, the energy consumed by a single reasoning is less than 5J.

S502a: The model management function module in the NMS sends reasoning capability query information to the model reasoning function module in the EMS. The reasoning capability query information is used to query (or obtain) the reasoning capability of the model reasoning function module in the EMS.

S502b: The model reasoning function module in the EMS sends reasoning capability information to the model management function module in the NMS.

Correspondingly, the model management function module in the NMS receives the reasoning capability information.

In one implementation, the above step S502a may also be omitted, that is, the model management function module in the NMS does not send reasoning capability query information to the model reasoning function module in the EMS, but the model reasoning function module in the EMS actively reports (i.e., sends) the reasoning capability information to the model management function module in the NMS.

The above-mentioned reasoning capability information may include: reasoning capability indication information, reasoning computing power (optional), and storage space (optional). The storage space may be the size of the storage space occupied by the model reasoning, or the storage address, etc. Among them, the reasoning capability indication information is used to indicate whether the model reasoning function in the EMS supports multi-model reasoning.

Exemplarily, if the corresponding value of the reasoning capability indication information is true (yes/correct), it means that multi-model reasoning is supported; if the corresponding value of the reasoning capability indication information is false (no/error), it means that multi-model reasoning is not supported; or the reasoning capability indication information is a specific value, if the value of the reasoning capability indication information is 1, it means that multi-model reasoning is supported, and if the value of the reasoning capability indication information is 0, it means that multi-model reasoning is not supported.

The inference computing power may refer to the computing power information available at the model inference function module, such as the available hardware resource information and the utilization rate of the hardware resources, wherein the hardware resources may include general computing power, such as the central processing unit (CPU), and high-performance computing power, such as the graphics processing unit (GPU), the neural network processing unit (NPU), etc. The hardware resource information may be the original hardware information, which may include the hardware type, the number of cores, the processing frequency, etc., or the quantified computing power, which may usually be measured by the number of floating-point operations per second (FLOPS) supported.

S503a: The model management function module in the NMS sends training capability query information to the model training function module in the EMS. Correspondingly, the model training function module in the EMS receives the training capability query information. The training capability query information is used to query (or obtain) the training capability of the model training function model in the EMS.

S503b: The model training function module in the EMS sends training capability information to the model management function module in the NMS.

Of course, the model management function module in the NMS may not send the training capability query information to the model training function module in the EMS, that is, the above step S503a is not executed, but the model training function module in the EMS actively sends the training capability query information to the model management function module in the NMS. The module reports (i.e. sends) its own training capability information.

Among them, the above-mentioned training capability information is used to notify (or indicate) whether the model training function model in the EMS supports multi-model training.

Exemplarily, the model training function module in the EMS sends training capability indication information to the model management function module in the NMS. If the corresponding value of the training capability indication information is true (yes or correct), it means that multi-model training is supported; if the corresponding value of the training capability indication information is false (no/error), it means that multi-model training is not supported; or the training capability indication information is a specific value. If the value of the training capability indication information is 1, it means that multi-model training is supported; if the value of the training capability indication information is 0, it means that multi-model training is not supported.

The above steps S501a-S501b, S502a-S502b, S503a-S503b are the query and reporting process of reasoning requirement information and capability information, which are optional steps, or can be completed offline. In addition, the embodiment of the present application does not specifically limit the order of executing the above-mentioned steps of querying and reporting reasoning requirement information (i.e., S501a-S501b), querying and reporting reasoning capability information (i.e., S502a-S502b), and querying and reporting training capability information (i.e., S503a-S503b).

Through the above steps S501a-S501b, S502a-S502b, S503a-S503b, the model management function module in the NMS can determine the reasoning requirements of the model reasoning function module in the EMS, the training capability of the model training function module in the EMS, and the reasoning capability of the model reasoning function module in the EMS, so as to further perform the following steps:

S504: The model management function module in the NMS sends a model training request to the model training function module in the EMS. Correspondingly, the model training function module in the EMS receives the model training request.

The model training request includes: model identification (or inference type), multi-model training instructions, and multi-model training strategy (optional).

The model identifier may also be expressed as an inference type. The multi-model training indication is used to indicate whether to perform multi-model training. The multi-model training strategy is used to indicate the training method, which may include a data processing strategy and a training algorithm indication.

In the above, the data processing strategy may include input data sampling and feature sampling. Among them, input data sampling indicates sampling the original data to form multiple sub-data sets, each of which is used to train a sub-model. Feature sampling indicates sampling the features of the original data, and data with different features can form multiple different sub-data sets, each of which is used to train a sub-model.

The training algorithm indication may include: the number of sub-models, the type of model, and the hyper-parameter configuration. The number of sub-models indicates the number of sub-models that constitute the multi-model; the type of model indicates the model type of different sub-models, such as random forest model and convolutional neural network model; the hyper-parameter configuration indicates the hyper-parameters of model training, such as the number of layers, number of iterations, and learning rate of the neural network model.

S505: The model training function module in the EMS performs multi-model training according to the model training request to obtain a first model (multi-model).

If the model training request does not include a multi-model training strategy, the model training function module in the EMS can determine the multi-model training strategy by itself based on the obtained reasoning requirement information. The specific content of the multi-model training strategy can refer to the content of the multi-model training strategy in the above step S504, which will not be introduced in detail here.

S506: The model training function module in the EMS sends a model training report to the model management function module in the NMS. Correspondingly, the model management function module in the NMS receives the model training report.

The model training report may include: model information of the first model, indication information of multiple models, a list of sub-models of the first model (i.e., information of multiple sub-models), an aggregation method (also called a combination method), a weight factor (optional), the performance of the first model, the size of the first model, the computational complexity of the first model, the inference speed of the first model, and the inference energy consumption of the first model.

Exemplarily, the model information of the first model may be information used to identify the first model, such as a name or a unique identifier.

Exemplarily, the indication information of the multiple models is used to indicate that the first model trained by the model training function module in the EMS is a multiple model. The list of sub-models of the first model contains a series of information about the multiple sub-models constituting the first model, and the information of each sub-model includes: identification information of the sub-model, the level of the sub-model (also referred to as the category of the sub-model), the performance of the sub-model, and the performance constraint of the sub-model. The identification information of the sub-model can be a unique identifier of the sub-model or a storage address of the sub-model. The level of the sub-model can include a first-level sub-model (equivalent to the first-level sub-model in the present application) and a second-level sub-model (equivalent to the second-level sub-model in the present application, also referred to as an aggregate model), and the second-level sub-model is used to aggregate the reasoning information of the first-level sub-model.

In the model training report, if the list of sub-models of the first model contains only multiple first-level sub-models, then the model training report also includes an aggregation method, which can be but is not limited to voting method, simple average method, weighted average method, linear mixing method. When the aggregation method included in the model training report is weighted average method or linear mixing method, then the model training report should also include the weight factor corresponding to each sub-model. If the list of sub-models of the first model contains multiple first-level sub-models and one second-level sub-model, then the model training report may not include the aggregation method (also called the combination method) and the weight factor.

S507: The model management function module in the NMS determines the model actually deployed.

That is, the model management function module in the NMS adjusts the first model and determines the sub-model to be actually deployed based on the reasoning requirement information and reasoning capability information from the EMS and the first model information in the model training report.

In this first embodiment, the model management function module in the NMS can adjust the number and aggregation method of sub-models actually deployed of the first model based on the reasoning requirement information and reasoning capability information of the model reasoning function module in the EMS and the model information in the training report.

For example, the model management function module in the NMS reduces the number of sub-models actually deployed in the first model according to the performance of each sub-model and the performance constraints of each sub-model.

S508a: The model management function module in the NMS sends model deployment request information to the model reasoning function module in the EMS.

Correspondingly, the model reasoning function module in the EMS receives the model deployment request information, which is used to request the deployment of the sub-model of the first model. The model deployment request information includes the model information of the first model, the indication information of multiple models, the list of sub-models of the first model (i.e., the information of multiple sub-models), the aggregation method (also called the combination method), and the weight factor (optional).

Then, the model reasoning function module in the EMS actually deploys the sub-model of the first model trained by the model training function module in the EMS based on the information of the sub-model of the first model contained in the model deployment request.

S508b: The model reasoning function module in the EMS sends a model deployment response message to the model management function module in the NMS. Accordingly, the model management function module in the NMS receives the model deployment response message to determine (or know) that the model reasoning function module in the EMS has completed the deployment of the sub-model of the first model.

S509: The model reasoning function module in the EMS performs multi-model reasoning based on the first model to obtain a reasoning result.

That is, the model reasoning function module in the EMS inputs the data to be inferred (i.e., input data) into the sub-models actually deployed in the first model, respectively, to obtain the reasoning results of the corresponding sub-models. Furthermore, the model reasoning function model can combine the reasoning results of each sub-model based on an aggregation method to obtain the final reasoning result. Alternatively, the model reasoning function model can input the reasoning results obtained by each first-level sub-model of the first model into the second-level sub-model, and output the combined final reasoning result.

For example, the model reasoning function model in the EMS can determine the multiple first-level sub-models according to the information of the multiple first-level sub-models of the first model included in the model reasoning request information, use the multiple first-level sub-models to perform reasoning respectively, and then obtain the reasoning results of the multiple first-level sub-models based on the storage addresses of the multiple first-level sub-models, and finally use the aggregation method to aggregate (or combine) the reasoning results of the multiple first-level sub-models to obtain the final reasoning result.

Alternatively, the model reasoning function model in the EMS can determine the multiple first-level sub-models based on the information of the multiple first-level sub-models of the first model included in the model reasoning request information, use the multiple first-level sub-models to perform reasoning respectively, and then obtain the reasoning results of the multiple first-level sub-models based on the storage addresses of the multiple first-level sub-models, and finally use the second-level sub-model to aggregate (or combine) the reasoning results of the multiple first-level sub-models to obtain the final reasoning result.

In this embodiment 1, the model training function module and the model reasoning function module in the EMS can correspondingly feedback the training capability (whether to support the training of multiple models), the reasoning capability (whether to support the reasoning of multiple models) and the reasoning requirement information of the multiple models to the model management function module in the NMS; in the EMS, the model management function module adds the training instructions of multiple models and the training strategy of multiple models to the model training request sent by the model training function module, and then the model training function module can generate the first model (i.e., multiple models) based on the model training request, send the information of the first model to the model management function module, add the indication information of multiple models and the information of the sub-models of the first model to the information of the first model, and then the model management module adjusts the number of sub-models of the first model based on the information of the first model, and the reasoning requirements and reasoning capabilities of the model reasoning function module to determine the first model actually deployed. Therefore, through this embodiment 1, it can be known that the scheme of the present application supports the reasoning function module of the management domain to perform multi-model combined reasoning, and supports the use of the most appropriate multi-model according to the reasoning requirements, which can effectively improve the reasoning (or analysis) effect of the model.

Embodiment 2:

In this second embodiment, the solution of the present application is applied to the deployment architecture shown in FIG. 3B above, and the model training and deployment process of the OAM (NMS/EMS) domain is enhanced to support reasoning and analysis of the RAN domain based on the combination of multiple learner models. This second embodiment is similar to the steps of the first embodiment above. The difference of this second embodiment is that the model reasoning function module is located in the RAN/gNB, that is, the first communication device in the solution of the present application is a network element management system device (EMS for short) including a model training function module, the second communication device in the solution of the present application is a RAN/gNB including a model reasoning function module, and the third communication device in the solution of the present application is a network management system device (NMS for short) including a model management function module. In addition, the model reasoning function module in the RAN/gNB and the model management function module in the NMS device are similar. The modules can interact directly with each other or forward through EMS. The solution of the embodiment of the present application can be implemented when the model reasoning function model in the RAN/gNB supports multi-model reasoning. Referring to Figure 6, the specific process of the second embodiment is as follows:

S601a: The model management function module in the NMS sends reasoning requirement query information to the model reasoning function module in the RAN/gNB.

In an optional implementation, the model management function module in the NMS can forward the reasoning requirement query information to the model reasoning function module in the RAN/gNB through the EMS. The reasoning capability query information is used to query (or obtain) the reasoning capability of the model reasoning function module in the RAN/gNB.

S601b: The model reasoning function module in the RAN/gNB sends reasoning requirement information to the model management function module in the NMS.

In an optional implementation, the model reasoning function module in the RAN/gNB can forward the reasoning requirement information to the model management function module in the NMS through the EMS. The reasoning capability information may include: reasoning capability indication information, reasoning computing power (optional), and storage space (optional).

The content of the reasoning capability information in step S601b can refer to the above step S501b, which will not be described in detail here.

S602a: The model management function module in the NMS sends reasoning capability query information to the model reasoning function module in the RAN/gNB.

In an optional implementation, the model management function module in the NMS may forward the reasoning capability query information to the model reasoning function module in the RAN/gNB through the EMS. The reasoning capability query information is used to query (or obtain) the reasoning capability of the model reasoning function module in the RAN/gNB.

S602b: The model reasoning function module in the RAN/gNB sends reasoning capability information to the model management function module in the NMS.

In an optional implementation, the model reasoning function module in the RAN/gNB may forward the reasoning capability information to the model management function module in the NMS through the EMS. The reasoning capability information may include: reasoning capability indication information, reasoning computing power (optional), and storage space (optional).

The content of the reasoning capability information in step S602b can refer to the specific description in the above step S502b, which will not be repeated here.

S603a: The model management function module in the NMS sends training capability query information to the model training function module in the EMS. The specific description of step S603a can be found in the above step S503a, which will not be repeated here.

S603b: The model training function module in the EMS sends the training capability information to the model management function module in the NMS. The specific description of step S603b can be found in the above step S503b, which will not be repeated here.

The above steps S601a-S601b, S602a-S602b, S603a-S603b are the query and reporting process of reasoning requirement information and capability information, which are optional steps, or can be completed offline. In addition, the embodiment of the present application does not specifically limit the order of executing the above-mentioned steps of querying and reporting reasoning requirement information (i.e., S601a-S601b), querying and reporting reasoning capability information (i.e., S602a-S602b), and querying and reporting training capability information (i.e., S603a-S603b).

S604: The model management function module in the NMS sends a model training request to the model training function module in the EMS. The specific description of step S604 can be found in the above step S504, which will not be repeated here.

S605: The model training function module in the EMS performs multi-model training according to the model training request to obtain a first model (multi-model). The specific description of step S605 can be found in the above step S505, which will not be repeated here.

S606: The model training function module in the EMS sends a model training report to the model management function module in the NMS. The specific description of step S606 can be found in the above step S506, which will not be repeated here.

S607: The model management function module in the NMS determines the model to be actually deployed. That is, the model management function module in the NMS adjusts the first model and determines the sub-model to be actually deployed based on the reasoning requirement information and reasoning capability information, as well as the first model information in the training report. The specific description of step S607 can be found in the above step S507, which will not be repeated here.

S608a: The model management function module in the NMS sends model deployment request information to the model reasoning function module in the RAN/gNB.

In an optional implementation, the model management function module in the NMS can forward the model deployment request information to the model reasoning function module in the RAN/gNB through the EMS. The content of the model deployment request information in step S608a can refer to the specific description in the above step S508a, which will not be repeated here.

S608b: The model reasoning function module in the RAN/gNB sends model deployment response information to the model management function module in the NMS.

In an optional implementation, the model reasoning function module in the RAN/gNB can send the model deployment response information to the model management function module in the NMS through the EMS. The content of the model deployment response information in step S608b can refer to the description of the model deployment response information in the above step S508b.

S609: The model reasoning function module in the RAN/gNB performs reasoning based on the first model to obtain a reasoning result.

Exemplarily, the model reasoning function module in the RAN/gNB inputs the data to be reasoned (i.e., input data) into the sub-models actually deployed in the first model, respectively, to obtain the reasoning results of the corresponding sub-models. The reasoning results of each sub-model are combined to obtain the final reasoning result. Alternatively, the model reasoning function model inputs the reasoning results obtained by each first-level sub-model of the first model into the second-level sub-model, and outputs the combined final reasoning result.

The model reasoning function module in the RAN/gNB executes step S609. The specific description of the model reasoning function module in the above-mentioned EMS executing step S509 can be referred to, and will not be repeated here.

Compared with the above-mentioned embodiment 1, through this embodiment 2, it can be seen that the solution of the present application supports the reasoning function module of the RAN domain to perform reasoning and combination of multiple models, and supports the use of the most appropriate model for reasoning according to the reasoning requirements, which can effectively improve the reasoning (or analysis) effect of the model.

Embodiment three:

In this third embodiment, the solution of the present application is applied to the deployment architecture shown in FIG. 3C above, and the model discovery and subscription process of NWDAF is enhanced to support NWDAF reasoning (or analysis) based on multi-model combination, thereby improving the reasoning (or analysis) effect of NWDAF. The first communication device in the solution of the present application is the first NWDAF network element including the model training function module in FIG. 3C, and the second communication device in the solution of the present application is the second NWDAF network element including the model reasoning function module in FIG. 3C. Referring to FIG. 7, the specific process of the third embodiment is as follows:

S701: The first NWDAF network element sends NF registration request information to the NRF network element. Correspondingly, the NRF network element receives the NF registration request information.

The NF registration request information may include an inference identifier (also called an analysis identifier), an indication of the ability to support multi-model training, inference performance, inference speed (optional), and inference power consumption (optional).

S702: The NRF network element sends the response information of the NF registration request to the first NWDAF network element. Correspondingly, the first NWDAF network element receives the response information of the NF registration request.

Through the above steps S701-S702, the model training function module of the first NWDAF network element reports its own model training capability information to the NRF network element.

In the embodiment of the present application, the first NWDAF network element is taken as an example to introduce its reporting of its own model training capability information to the NRF network element. In practice, there may be multiple NWDAF network elements including model training functions that report their own model training capability information to the NRF network element. Each NWDAF network element including model training functions can report its own model training capability information to the NRF network element by referring to the above steps S701-S702.

S703: The NWDAF consumer sends analysis subscription request information to the second NWDAF network element. Correspondingly, the second NWDAF network element receives the analysis subscription request information.

S704: The second NWDAF network element sends the response information of the analysis subscription request to the NWDAF consumer. Correspondingly, the NWDAF consumer receives the response information of the analysis subscription request.

S705: The second NWDAF network element sends NF discovery request information to the NRF network element. Correspondingly, the NRF network element receives the NF discovery request information, and the NF discovery request information (equivalent to the reasoning requirement information in the above-mentioned solution of the present application) may include: NF type, reasoning identifier, reasoning performance requirement, multi-model training capability indication information, reasoning speed requirement, and reasoning power consumption requirement.

S706: The NRF network element sends a response message of the NF discovery request to the first NWDAF network element. Correspondingly, the first NWDAF network element receives the response message of the NF discovery request. The response message of the NF discovery request includes the address of the NWDAF network element that has the model training function and supports multi-model training.

Since in the third embodiment, the NWDAF network element that has the model training function and supports multi-model training takes the first NWDAF network element as an example, the response information of the NF discovery request includes the address of the first NWDAF network element.

S707: The second NWDAF network element sends model subscription request information to the first NWDAF network element. Correspondingly, the first NWDAF network element receives the model subscription request information.

The model subscription request information (equivalent to the model request information in the above-mentioned solution of the present application) includes an inference identifier, inference performance requirements, indication information for requesting multi-model inference (the indication information is used to request multiple models), inference speed requirements, and inference power consumption requirements (these are equivalent to the inference requirement information in the above-mentioned solution of the present application).

S708: The first NWDAF network element sends response information of the model subscription request to the second NWDAF network element. Correspondingly, the second NWDAF network element receives the response information of the model subscription request.

S709: The first NWDAF network element performs multi-model training according to the model reading request information to obtain a first model (ie, multi-model).

In the third embodiment, step S709 is an optional step, that is, step S709 may be executed or not executed. When step S709 is not executed, the first NWDAF network element may, based on the inference requirement information included in the model subscription request information, Directly select a suitable multi-model (ie, the first model) from at least one trained multi-model.

S710: The first NWDAF network element sends model notification information to the second NWDAF network element. Correspondingly, the second NWDAF network element receives the model notification information, which includes an inference identifier, indication information of the first model, a sub-model list of the first model (i.e., information of multiple sub-models), an aggregation method (also referred to as a combination method), a weight factor (optional), and performance information of the first model.

The indication information of the first model is used to indicate that the first model is a multi-model. The information of each sub-model may include but is not limited to: identification information of the sub-model (such as a unique identifier or storage address information of the sub-model), the level of the sub-model (also referred to as the category of the sub-model), the performance of the sub-model, and the performance constraint of the sub-model.

For the specific explanation of the above-mentioned aggregation method, the performance of the first model, and the relevant information of the sub-model, please refer to the specific introduction in the above-mentioned embodiment 1, which will not be repeated here.

S711: The second NWDAF network element performs multi-model reasoning (or analysis) based on the model notification information to obtain a reasoning (or analysis) result.

Exemplarily, the step S711 may be specifically executed with reference to the manner in which the model reasoning function module in the above step S509 or S609 performs reasoning based on the first model to obtain a reasoning result, which will not be described in detail here.

S712: The second NWDAF network element sends notification information of the reasoning result to the NWDAF consumer, where the notification information of the reasoning result includes the reasoning identifier, the reasoning result or the analysis result. Correspondingly, the NWDAF consumer receives the notification information of the reasoning result.

In this third embodiment, the first NWDAF network element including the model training function module and the second NWDAF network element including the model reasoning function module can report the corresponding training capabilities and model capabilities, and reasoning requirement information to the NRF network element, and the first NWDAF network element including the model training function module and the second NWDAF network element including the model reasoning function module can exchange reasoning requirements and multi-model indication information. Therefore, the solution of this third embodiment supports NWDAF to select (or train) appropriate multi-models according to reasoning requirements, and supports NWDAF to perform multi-model combined reasoning, thereby effectively improving the reasoning (or analysis) effect of NWDAF.

Embodiment 4:

In this fourth embodiment, the solution of this application is applied to the deployment architecture shown in FIG. 3D above, by enhancing the RAN model deployment and switching process to support intelligent reasoning (or analysis) based on multi-model reasoning. The first communication device in this application solution can be a base station gNB (i.e., source gNB\target gNB) including a model training function module in FIG. 3D, and the second communication device in this application solution is UE1 including a model reasoning function model in FIG. 3D. Referring to FIG. 8, the specific process of this fourth embodiment is as follows:

S801a: The target gNB sends the AI capability information of the target gNB to the source gNB. Correspondingly, the source gNB receives the AI capability information of the target gNB.

S801b: The source gNB sends the AI capability information of the source gNB to the target gNB. Correspondingly, the target gNB receives the AI capability information of the source gNB.

The above steps S801a and S801b take a source gNB and a target gNB as an example to introduce the process of exchanging AI capability information between the source gNB and the target gNB. In practical applications, the gNB that exchanges AI capability information with the source gNB is not limited to the target gNB. Exemplarily, after the target gNB and the source gNB go online, they can exchange AI capability information through the Xn interface. The AI capability information may include: AI switch, training capability indication information (i.e., supporting multi-model training).

S802a: The source gNB sends reasoning capability query information to UE1.

Exemplarily, taking UE1 as an example, after UE1 accesses the source gNB, the source gNB sends reasoning capability query information to UE1, and the reasoning capability query information is used to query (or obtain) the reasoning capability information of UE1.

S802b: UE1 sends reasoning capability information to the source gNB.

Exemplarily, the UE1 may send reasoning capability information to the source gNB through the Uu interface. Alternatively, after the UE1 accesses the source gNB, the source gNB does not send reasoning capability query information to the UE1, but the UE1 actively reports its own reasoning capability information to the source gNB through the Uu interface.

Among them, the reasoning capability information of UE1 may include: AI switch, storage space size, reasoning computing power, multi-model reasoning indication information (i.e., indicating support for multi-model reasoning), and remaining power of UE1 (optional).

The embodiment of the present application does not specifically limit the time sequence of executing the process of exchanging AI capability information between the source gNB and the target gNB (i.e., the above steps S801a-S801b) and the process of UE1 reporting reasoning capability information to the source gNB (i.e., the above steps S802a-S802b).

S803: The source gNB performs multi-model training based on the reasoning capability information of UE1 to obtain a first model (i.e., a multi-model). This step S803 is an optional step, which may be executed or not.

In a possible implementation, the source gNB may also directly select a suitable multi-model (i.e., the first model) from at least one trained multi-model based on the reasoning capability information of the UE1.

S804: The source gNB sends notification information of the first model to UE1.

Exemplarily, the source gNB may send notification information of the first model to the UE1 through the Uu interface, and correspondingly, the UE1 receives the notification information of the first model.

The notification information of the first model may include: the identifier of the first model, indication information of the first model (i.e., used to indicate that the first model is a multi-model), a list of sub-models of the first model (i.e., information of multiple sub-models), an aggregation method (also called a combination method), a weight factor (optional), and the performance of the first model.

S805a: The source gNB may perform multi-model inference based on the information of the first model to obtain an inference result.

This step S805a is an optional step. If the first model is a bilateral model, the source gNB executes this step S805a; if the first model is a unilateral UE model, the source gNB does not execute this step S805a.

S805b: The UE1 may perform multi-model reasoning based on the information of the first model to obtain a reasoning result.

At this time, the source gNB also determines whether it is necessary to switch the base station (gNB) accessed by UE1 based on the received signal strength reported by UE1. For example, when the signal strength received by the source gNB from UE1 is lower than the set threshold, it is determined to trigger the switch.

When the source gNB determines to switch the gNB to which UE1 is to be connected, the following steps are performed:

S806: The source gNB sends an RRC connection reconfiguration message to UE1, which includes measurement configuration information.

Exemplarily, the source gNB may send the RRC connection reconfiguration message to the UE1 via the Uu interface. Correspondingly, the UE1 receives the RRC connection reconfiguration message via the Uu interface.

S807: UE1 performs measurement based on the measurement configuration information and obtains a measurement report of UE1.

S808: UE1 sends a measurement report of UE1 to the source gNB.

Exemplarily, the UE1 may send a measurement report of the UE1 to the source gNB via the Uu interface. Correspondingly, the source gNB receives the measurement report of the UE1 via the Uu interface.

S809: The source gNB determines the target gNB based on the measurement report of UE1 and the AI capability information of the neighboring station.

It can be understood that the target gNB generally has the ability to train multiple models. For example, gNB1, gNB2, and gNB3 refer to the above steps S801a and S801b to exchange AI capability information with the source gNB, so that the source gNB obtains the AI capability information of gNB1, the AI capability information of gNB2, and the AI capability information of gNB3; and then in this step, the source gNB can select a suitable gNB1 as the target gNB from the three base stations based on the AI capability information of the three base stations and the measurement report of UE1.

S810: The source gNB sends the handover request information of UE1 to the target gNB. Correspondingly, the target gNB receives the handover request information of UE1. The handover request information of UE1 includes the identifier of UE1, the indication information of the first multi-model (or the indication information of using the first model), the identifier of the first model, and the information of the sub-model of the first model.

S811: The target gNB performs multi-model training based on the handover request information of UE1 to obtain a second model (i.e., multi-model).

This step S811 is an optional step, that is, step S811 may be executed or not. If the target gNB does not execute step S811, the target gNB may directly select a multi-model (i.e., the second model) from at least one trained multi-model based on the handover request information of the UE1; or the target gNB may directly use the multi-model (i.e., the first model) of the source gNB.

S812: The target gNB and UE1 complete random access.

In step S812, the target gNB performs a random access process with the UE1 so that the UE1 successfully accesses the target gNB for communication. The random access process can be specifically implemented by referring to the existing random access method and will not be described in detail here.

S813: The target gNB sends notification information of the second model to UE1.

Exemplarily, the target gNB may send notification information of the second model to UE1 through the Uu interface, and correspondingly, the UE1 receives notification information of the second model through the Uu interface.

The notification information of the second model may include model information of the second model (such as the name, identifier, type, etc. of the second model), indication information of the second model (used to indicate that the second model is a multi-model), a list of multiple sub-models of the second model (i.e., information of multiple sub-models), aggregation method, weight factor (optional), and performance of the second model.

The list of each sub-model (or the information of each sub-model) may include: the level of the sub-model (also referred to as the category of the sub-model), the identification information of the sub-model (such as the storage address of the sub-model, the unique identifier).

The level of sub-models may include a first-level sub-model and a second-level sub-model, wherein the first-level sub-model is used for reasoning or analysis, and the second-level sub-model is used to aggregate (or combine) reasoning information of multiple first-level sub-models.

For detailed explanations of the above-mentioned aggregation method, the performance of the second model, and the related information of the sub-models of the second model, please refer to The detailed introduction of the aggregation method, the performance of the first model and the related information of the sub-models of the first model in the above-mentioned embodiment 1 will not be repeated here.

S814a: The target gNB performs multi-model inference based on the notification information of the second model to obtain an inference result.

That is, the target gNB uses the sub-models of the second model for reasoning based on the information of the second model in the notification information of the second model to obtain reasoning results of multiple sub-models, and combines them in an aggregation manner to obtain the final reasoning result; or uses the first-level sub-models of the second model for reasoning to obtain reasoning results of multiple first-level sub-models, and then uses the second-level sub-model to combine the reasoning results of these first-level sub-models to obtain the final reasoning result.

The step S814a is an optional step, and the step S814a may be performed or may not be performed.

S814b: UE1 performs multi-model inference based on the notification information of the second model to obtain an inference result.

The model reasoning function module in UE1 can perform reasoning using the multiple sub-models based on the information of the multiple sub-models of the second model in the notification information of the second model to obtain the reasoning information of the multiple sub-models, and then use an aggregation method to aggregate or combine the reasoning information of the multiple sub-models to obtain the reasoning information of the second model; or the model reasoning function module in UE1 can use the second sub-model to aggregate or combine the reasoning information of the multiple first-level sub-models to obtain the reasoning information of the second model.

The details of step S814b may also refer to the above-mentioned step S509 or S609 or S711, which will not be described in detail here.

Through the above-mentioned fourth embodiment, the multi-model training capability information interaction between new base stations and the multi-model training/inference capability interaction between the base station and the terminal are added, and the indication information of the multi-model is added in the model notification information. Through this fourth embodiment, it can be known that the present application scheme supports the UE to perform combined reasoning based on multiple models to improve the reasoning (or analysis) effect of the model. In addition, when the source gNB determines to trigger the switching of a new gNB, the source gNB will, based on the multi-model capability information of each gNB, preferably use a base station with multi-model training capability as the target gNB for switching. After the UE switches to access the target gNB, it can still perform combined reasoning based on multiple models to improve the reasoning (or analysis) effect of the model.

Embodiment five:

In this fifth embodiment, the solution of the present application is applied to the deployment architecture shown in FIG. 3E above, that is, a scenario in which joint training of a base station and a UE is performed in a bilateral model scenario. In this fifth embodiment, the first communication device and the second communication device in the solution of the present application may be a base station (gNB) or a UE (e.g., UE1) including a model training function model and a model reasoning function module. Referring to FIG. 9 , the specific process of the fifth embodiment is as follows:

S901a: The target gNB sends the AI capability information of the target gNB to the source gNB.

Exemplarily, after the target gNB and the source gNB come online, they can exchange AI capability information through the Xn interface. The AI capability information may include: AI switch, support for multi-model training capability indication (yes/no).

S901b: The source gNB sends the AI capability information of the source gNB to the target gNB.

Referring to the above steps S901a and S901b, the source gNB and multiple gNBs (including the target gNB) can exchange their respective AI capability information through the corresponding Xn interface. This embodiment 5 takes the source gNB and the target gNB as an example.

S902a: The source gNB sends capability query information to UE1.

Exemplarily, after UE1 accesses the source gNB, the source gNB may send capability query information to the source UE1 through the Uu interface, and the capability query information is used to query (or request) the capability information of the UE1. Correspondingly, the UE1 receives the capability query information through the Uu interface.

S902b: UE1 sends UE1’s capability information to the source gNB.

Exemplarily, UE1 may send the capability information of UE1 to the source gNB through the Uu interface, and correspondingly, the source gNB receives the capability information of UE1 through the Uu interface. Alternatively, the source gNB does not send the capability query information to UE1, but UE1 actively sends the capability information of UE1 to the source gNB.

Among them, the capability information of UE1 may include: AI switch, storage space, computing power, support for multi-model reasoning indication information, and remaining power (optional).

The embodiment of the present application does not specifically limit the order of executing the steps of exchanging AI capability information between the above-mentioned base stations (ie, S901a-S901b), and the steps of querying and reporting UE1 capability information (ie, S902a-S902b).

S903: The source gNB and UE1 negotiate a joint training strategy. The joint training strategy may include: multi-model training indication information, multi-model joint training mode, data processing strategy, number of sub-models, model type, and hyper-parameter configuration.

The multi-model training indication information is used to indicate the training of multiple models. The multi-model joint training mode can be one-to-one, many-to-one, one-to-many, or many-to-many; one-to-one means that the multiple models on the source gNB and the UE1 side are both regarded as an overall model, and the output of the gNB's overall model is used as the input of the UE1's overall model; many-to-one means that the multiple models on the UE1 side are regarded as an overall model, and the source gNB The multiple outputs of the multi-model on the source gNB side are used as the input of the overall model of UE1; one-to-many means that the multiple models on the source gNB side are used as an overall model, and the output of the overall model of the source gNB is used as the input of the multi-model of UE1; many-to-many means that the multiple outputs of the multi-model on the source gNB side are used as the input of the multi-model of UE1.

Among them, the data processing strategy can be input data sampling, feature sampling, etc. The input data sampling and feature sampling can be specifically described in the above embodiments and will not be described in detail here.

S904a: The source gNB performs multi-model training to obtain a multi-model of the source gNB.

S904b: UE1 performs multi-model training to obtain a multi-model of UE1.

The above steps S904a and S904b can be executed synchronously, and when the source gNB and the UE1 perform multi-model training respectively, the intermediate parameters of their respective multi-model training, such as gradients or intermediate inference results, are exchanged according to the joint training strategy in the above step S903.

S905a: The source gNB performs inference based on the multiple models and aggregation method of the source gNB to obtain an inference result.

Specifically, the source gNB uses its own trained sub-models to perform reasoning separately, and then uses the aggregation method to combine the reasoning results of the sub-models to obtain the final reasoning result.

S905b: UE1 performs reasoning based on the multiple models and aggregation method of UE1 to obtain a reasoning result.

Specifically, the UE1 uses each sub-model trained by itself to perform reasoning respectively, and then uses an aggregation method to combine the reasoning results of each sub-model to obtain a final reasoning result.

In one implementation, the above step S905b may be executed first, and then step S905a may be executed, that is, after UE1 obtains the final inference result using the multi-model trained by itself, the final inference result on the UE1 side is reported to the source gNB. The source gNB may use the final inference result on the UE1 side as the input of the multi-model of the source gNB to obtain the final inference result on the source gNB side.

When the source gNB determines that the base station to which UE1 accesses needs to be switched, the following steps are continued:

S906: The source gNB sends measurement configuration information to UE1. Correspondingly, UE1 receives the measurement configuration information.

S907: UE1 performs measurement based on the measurement configuration information and obtains a measurement report of UE1.

S908: UE1 sends a measurement report of UE1 to the source gNB. Correspondingly, the source gNB receives the measurement report of UE1.

S909: The source gNB selects the target gNB based on the measurement report of UE1 and the AI capability of the neighboring station.

The source gNB preferably selects a base station with multi-model training capability as the target gNB based on the measurement report of UE1 and the AI capability of the neighboring base station.

S910: The source gNB sends the handover request information of UE1 to the target gNB. Correspondingly, the target gNB receives the handover request information of UE1, and the UE handover request information includes the identification information of UE1 and the multi-model indication information, and the multi-model indication information is used to request the use of the multi-model.

Optionally, in step S910, the source gNB may also send its own multi-model and the usage information of the multi-model of the source gNB to the target gNB, then the target gNB and UE1 directly reuse the multi-model of the source gNB and the multi-model previously trained by UE1, and there is no need to perform multi-model training separately, that is, the following steps S913a and S913b are not executed. The handover request information of the UE1 sent by the source gNB to the target gNB also includes the identifier of the multi-model trained by the source gNB, the list of the multi-model, the aggregation method, and the weight factor (optional).

Optionally, in step S910, the source gNB may also send the joint training strategy previously negotiated with the UE1 to the target gNB. In this case, there is no need to repeatedly negotiate the joint training strategy between the target gNB and the UE1, that is, the following step S912 is not executed.

S911: The target gNB completes random access with UE1.

The target gNB performs a random access process with the UE1 so that the UE1 successfully accesses the target gNB for communication. The specific random access process is implemented with reference to the existing random access process and will not be described in detail here.

S912: The target gNB and UE1 negotiate a joint training strategy.

The joint training strategy may include: multi-model training instructions, multi-model joint training mode, data processing strategy, number of sub-models, model type, and hyper-parameter configuration.

Step 912 is an optional step. If in the above step S910, the switching request information of the UE1 includes the joint training strategy, then the target gNB and the UE1 do not need to renegotiate the joint training strategy.

S913a: UE1 performs multi-model training to obtain a multi-model of UE1.

S913b: The target gNB performs multi-model training to obtain a multi-model of the target gNB.

In the above, when the target gNB and UE1 perform multi-model training respectively, they exchange intermediate parameters of the multi-model training, such as gradients or intermediate inference results, according to the negotiated joint training strategy.

S914a: UE1 performs reasoning and combination based on each sub-model of its own multi-model to obtain a reasoning result.

The UE1 uses each sub-model of the multi-model trained by itself to perform reasoning respectively, and then uses an aggregation method to combine the reasoning results of each sub-model to obtain a final reasoning result.

S914b: The target gNB performs inference and combination based on each sub-model of its own multi-model to obtain an inference result.

The target gNB uses each sub-model of its own trained multi-model (or each sub-model of the source gNB's multi-model) to perform inference respectively, and then uses the aggregation method to combine the inference results of each sub-model to obtain the final inference result.

In one embodiment, in the above step S914a, after UE1 obtains the final inference result using its own trained multi-model, it also reports the final inference result on the UE1 side to the target gNB. The target gNB can use the final inference result on the UE1 side as the input of the multi-model of the target gNB (or the multi-model of the source gNB) to obtain the final inference result on the target gNB side.

This fifth embodiment supports multi-model combined reasoning in the scenario where the base station and UE jointly train a bilateral model, which can improve the reasoning (or analysis) effect of the model. In addition, when switching the base station accessed by the UE, a base station with multi-model training capability is preferred, and the UE can still perform combined reasoning based on multiple models after switching.

Embodiment six:

In this embodiment six, another multi-model usage scenario is mainly aimed at, that is, the model training function can encapsulate multiple models into a large model, the model reasoning function can be unaware of the internal structure of the large model (i.e., multiple models), and the model selection and deployment process is added. This embodiment six is introduced with a general logical architecture, as shown in Figure 10, the specific process of this embodiment six is as follows:

S1001a: The model management function sends reasoning requirement query information to the model reasoning function.

In an optional implementation, the model management function forwards the reasoning requirement query information to the model reasoning function through the model training function.

This step S1001a can be cross-referenced with the above-mentioned step S501a or S601a.

S1001b: The model reasoning function sends reasoning requirement information to the model management function.

In an optional implementation, the model reasoning function forwards the reasoning requirement information to the model management function via the model training function.

In another optional implementation, the model reasoning function proactively reports (ie, sends) the reasoning requirement information to the model management function.

The reasoning requirement information includes: reasoning type requirement, reasoning accuracy requirement, reasoning speed requirement, and reasoning energy consumption requirement. Among them, the reasoning accuracy requirement, reasoning speed requirement, and reasoning energy consumption requirement can also be collectively referred to as reasoning performance requirements. The reasoning speed requirement can also be called the reasoning latency requirement, which indicates the requirement for reasoning execution time, for example: a single reasoning execution time is less than 1s; the reasoning energy consumption requirement indicates the requirement for reasoning energy consumption, for example: a single reasoning consumes less than 5J of energy.

The specific description of step S1001b may refer to the above step S501b or S601b, which will not be repeated here.

S1002a: The model management function sends reasoning capability query information to the model reasoning function.

In an optional implementation, the model management function forwards the reasoning capability query information to the model reasoning function via the model training function.

The step S1002a can be cross-referenced with the above-mentioned step S502a or S602a.

S1002b: The model reasoning function sends reasoning capability information to the model management function.

In an optional implementation, the model reasoning function forwards the reasoning capability information to the model management function via the model training function.

In another optional implementation, the model reasoning function proactively reports (ie, sends) the reasoning capability information to the model management function.

The reasoning capability information includes: reasoning computing power (optional), storage space (optional), power, etc. The reasoning computing power indicates the computing power information available at the reasoning function, including available hardware resource information and hardware resource utilization. The hardware resource information can be the original hardware information, including hardware type, number of cores, processing frequency, etc., or it can be the quantified computing power.

The specific description of step S1002b may refer to the above step S502b or S602b, which will not be repeated here.

S1003a: The model management function sends training capability query information to the model training function. This step S1003a can be cross-referenced with the above-mentioned step S503a or S603a.

S1003b: The model training function sends training capability information to the model management function. Exemplarily, the model training function can actively report the training capability information to the model management function. The training capability information includes training computing power and the upper limit of model accuracy that can be achieved.

The step S1003b may be cross-referenced with the above-mentioned step S503b or S603b.

S1004: The model management function sends model training request information to the model training function.

The model training request information includes: model identification/inference type, and model training strategy information.

The training strategy information of the model is determined based on the reasoning requirement information, the reasoning capability information and the training capability information; the training strategy information of the model is used to indicate the training method, which may include: multiple model training instruction information, data processing strategy, training algorithm instruction, etc. Alternatively, the model management function may also send the original reasoning requirement information and reasoning capability information to the model training function, and the model training function itself determines the model training strategy based on the reasoning requirement information, reasoning capability information, and training capability information.

The specific description of step S1004 may refer to the above-mentioned S504 or step S604, which will not be repeated here.

S1005: The model training function performs model training according to the model training request information to obtain a first model (ie, a multi-model).

The step S1005 may be described in detail with reference to step S505 or S605, and will not be repeated here.

S1006: The model training function sends a model training report to the model management function.

The model training report includes: identification information of the first model, accuracy of the model, accuracy constraints, size of the first model, inference computing power of the first model, inference speed of the first model, and inference energy consumption of the first model.

The step S1006 may be described in detail with reference to step S506 or S606, and will not be repeated here.

S1007: The model management function determines the model actually deployed.

That is, the model management function can determine the actually deployed model based on the original reasoning requirement information, reasoning capability information and the model training report.

For example, if the model management function in step S1004 also instructs the model training function to train multiple specified models (which may be multiple models similar to the first model), then in step S1006, the model training report fed back by the model training function may include multiple models with different performances, that is, multiple multiple models similar to the first model, but the performance of each model is different; the model management function can determine the actually deployed model (for example, the first model) from the multiple models with different performances based on the reasoning requirement information, reasoning capability information and the model training report.

S1008a: The model management function sends model deployment request information to the model reasoning function.

Correspondingly, the model reasoning function receives the model deployment request information, and the model deployment request information includes the actually deployed model information (such as the first model information).

Exemplarily, the model deployment request information includes: identification information of the first model actually deployed. Optionally, the model deployment request information may also include other information of the first model, such as the accuracy of the first model, the accuracy constraint of the first model, the size of the first model, the reasoning computing power of the first model, the reasoning speed of the first model, and the reasoning energy consumption of the first model.

S1008b: The model reasoning function sends model deployment response information to the model management function.

Correspondingly, the model management function receives the model deployment response information to determine (or know) that the model reasoning function has completed the model deployment.

S1009: The model reasoning function performs reasoning based on the actually deployed model to obtain a reasoning result.

Exemplarily, the model reasoning function uses the first model to perform reasoning based on the first model information (i.e., the model information actually deployed) to obtain a reasoning result. For example, the model reasoning function inputs the information to be reasoned into the first model and outputs a reasoning result.

In the sixth embodiment, a process of querying/reporting reasoning requirements is added between the model management function and the model reasoning function. The reasoning requirements include reasoning speed requirements, reasoning energy consumption requirements, etc. A process of querying/reporting reasoning capabilities is also added between the model management function and the model reasoning function. The reasoning capabilities include reasoning computing power, storage space, power, etc., and a process of querying/reporting training capabilities is added between the model management function and the model training function. The model training report sent by the model training function to the model management function includes relevant information of the model, such as model size, model reasoning computing power, model reasoning speed, model reasoning energy consumption, etc. The model management function can determine the model training strategy based on the reasoning requirements, reasoning capabilities, and training capabilities. The model management function can also determine the actual deployed model based on the reasoning requirements, reasoning capabilities, and the information of the model in the model training report. Therefore, in the sixth embodiment, querying/reporting reasoning requirements, reasoning capabilities, and training capabilities are added, and the trained model can be determined based on the reasoning requirements, reasoning capabilities, and training capabilities, and the best model for actual deployment can be determined based on the information of the reasoning requirements, reasoning capabilities, and actual models, thereby improving the reasoning (or analysis) effect of the model.

Embodiment seven:

The application scenario of this seventh embodiment is similar to that of the sixth embodiment, except that there is no model management function in the seventh embodiment, and only the interaction between the model training function and the model reasoning function is involved. The seventh embodiment is described in a general logical architecture, and specifically, it can be applied to the deployment architecture 3C-3E. Referring to FIG11 , the specific process of the seventh embodiment is as follows:

S1101: The model inference function sends model training request information to the model training function.

Correspondingly, the model training function receives the model training request information of the model inference function, and the model training request information includes: Model identification/inference type, inference accuracy requirement, inference speed requirement, inference energy consumption requirement, and request for multiple model instructions.

Among them, the value of requesting multiple models indication information is yes/no, to indicate whether it is necessary to provide multiple models that meet the reasoning requirements. When the requesting multiple models indication information indicates yes (that is, indicating that multiple models that meet the reasoning requirements need to be provided), it can further indicate the number of models required, for example, indicating that 5 models are required.

S1102: The model training function performs model training according to the request of the model inference function to obtain a first model (ie, a multi-model).

This step S1102 is an optional step. If the first model is a multi-model that has been trained in advance by the model training function, then step S1102 may not be performed.

S1103: The model training function sends the first model information to the model reasoning function.

Exemplarily, the first model information includes identification information of the first model (such as name, type), accuracy of the first model, accuracy constraint, size of the first model, inference computing power of the first model, inference speed of the first model, and inference energy consumption of the first model.

If the above step S1101 indicates that multiple models are required (which may be multiple models similar to the first model), then the first model information in step S1103 is a list containing multiple models with different performances, that is, it contains multiple model information similar to the first model, but the performance of each model is different.

S1104: The model reasoning function selects a suitable model based on the reasoning requirement information, the reasoning capability information and the first model information.

If the first model information returned in step S1103 is a list containing multiple similar first model information, then step S1104 is executed. Exemplarily, the model reasoning function selects a suitable model from the multiple models similar to the first model based on the reasoning requirement information, the reasoning capability information and the list.

S1105: The model reasoning function performs model reasoning to obtain reasoning results.

The model reasoning function may use the first model to perform reasoning to obtain a reasoning result, or the model reasoning function may use the model selected in the above step S1104 to perform reasoning to obtain a reasoning result.

In this embodiment seven, the inference speed requirement, inference energy consumption requirement, and request for multiple model indication information are added to the model training request information (or model request information) sent by the model inference function; the model training function can determine the returned model information according to the inference requirements, and add other information of the model to the model information, such as the size of the model, the model inference computing power, the model inference speed, and the model inference energy consumption; the model inference function can determine the model actually used based on the inference requirements, the inference ability, and the actual model information. This embodiment seven supports the determination of model information according to the inference requirements, and supports the selection of the most appropriate model for inference (or analysis) based on the inference requirements, the inference ability, and the actual model information, thereby improving the inference (or analysis) effect of the model.

The communication device provided in the embodiment of the present application is described below.

Based on the same technical concept, an embodiment of the present application provides a communication device, which can be used to perform the operation performed by the first communication device in the above method embodiment. The communication device can also be a first communication device, a processor of the first communication device, or a chip. The device includes a module or unit corresponding to the method/operation/step/action described by the first communication device in the above embodiment, and the module or unit can be a hardware circuit, or software, or a hardware circuit combined with software. The communication device can have a structure as shown in Figure 12.

As shown in FIG. 12 , the communication device 1200 may include a communication unit 1201 (also referred to as a transceiver unit) and a processing unit 1202. The communication unit 1201 is equivalent to a communication module (or a transceiver module), and the processing unit 1202 is equivalent to a processing module. The processing unit 1202 may be used to call the communication unit 1201 to perform a receiving and/or sending function, and the communication unit 1201 may implement a corresponding communication function. Specifically, the communication unit 1201 may include a receiving unit and/or a sending unit. The receiving unit may be used to receive information and/or data, and the sending unit may be used to send information and/or data. The communication unit 1201 may also be referred to as a communication interface or a transceiver module.

Optionally, the communication device 1200 may further include a storage unit 1203, which is equivalent to a storage module and can be used to store instructions and/or data. The processing unit 1202 can read the instructions and/or data in the storage module so that the communication device implements the aforementioned method embodiment.

The communication device 1200 can be used to perform the actions performed by the first communication device in the above method embodiment. The communication device 1200 can be the first communication device or a component that can be configured in the first communication device. The communication unit 1201 is used to perform the sending-related operations on the first communication device side in the above method embodiment, and the processing unit 1202 is used to perform the processing-related operations on the first communication device side in the above method embodiment.

Optionally, the communication unit 1201 may include a sending unit and a receiving unit. The sending unit is used to perform the sending operation in the above method embodiment. The receiving unit is used to perform the receiving operation in the above method embodiment.

It should be noted that the communication device 1200 may include a sending unit but not a receiving unit. Alternatively, the communication device 1200 may include a receiving unit but not a sending unit. Specifically, it may depend on whether the above solution executed by the communication device 1200 includes a sending action and a receiving action.

As an example, the communication device 1200 is used to execute the actions executed by the first communication device in the embodiment shown in FIG. 4A or FIG. 4B above.

For example, the communication unit 1201 is used to receive model request information, and the model request information includes reasoning requirement information; the processing unit 1202 is used to determine the first model according to the model request information, and the first model is a multi-model; the communication unit 1201 is also used to send first information, and the first information includes information of the first model.

It should be understood that the specific process of each unit executing the above corresponding process has been described in detail in the above method embodiment, and for the sake of brevity, it will not be repeated here.

The processing unit 1202 in the above embodiment may be implemented by at least one processor or processor-related circuits. The communication unit 1201 may be implemented by a transceiver or transceiver-related circuits. The storage unit may be implemented by at least one memory.

Based on the same technical concept, an embodiment of the present application provides a communication device, which can be used to perform the operation performed by the second communication device in the above method embodiment. The communication device can also be a second communication device, a processor of the second communication device, or a chip. The device includes a module or unit corresponding to the method/operation/step/action described by the second communication device in the above embodiment, and the module or unit can be a hardware circuit, or software, or a hardware circuit combined with software. The communication device can also have a structure as shown in Figure 12.

The communication device 1200 can be used to perform the actions performed by the second communication device in the above method embodiment. The communication device 1200 can be a first communication device or a component that can be configured in the second communication device. The communication unit 1201 is used to perform the sending-related operations on the second communication device side in the above method embodiment, and the processing unit 1202 is used to perform the processing-related operations on the second communication device side in the above method embodiment.

As an example, the communication device 1200 is used to execute the actions executed by the second communication device in the embodiment shown in FIG. 4A or FIG. 4B above.

For example, the communication unit 1201 is used to receive second information, which includes information of a first model, where the first model is determined based on reasoning requirement information, and the first model is a multi-model; the processing unit 1202 is used to obtain reasoning information of the first model based on the information of the first model.

Based on the same technical concept, an embodiment of the present application provides a communication device, which can be used to perform the operations performed by the third communication device in the above method embodiment. The communication device can also be a third communication device, a processor of the third communication device, or a chip. The device includes a module or unit corresponding to the method/operation/step/action described by the third communication device in the above embodiment. The module or unit can be a hardware circuit, or software, or a combination of a hardware circuit and software. The communication device can also have the following The structure shown in Figure 12.

As shown in FIG. 12 , the communication device 1200 may include a processing unit 1202, and optionally, a communication unit 1201. The communication unit 1201 is equivalent to a transceiver module, and the processing unit 1202 is equivalent to a processing module. The processing unit 1202 may be used to call the communication unit 1201 to perform a receiving and/or sending function, and the communication unit 1201 may implement a corresponding communication function. Specifically, the communication unit 1201 may include a receiving unit and/or a sending unit. The receiving unit may be used to receive information and/or data, and the sending unit may be used to send information and/or data. The communication unit 1201 may also be called a communication interface or a transceiver module.

The communication device 1200 may be used to perform the actions performed by the third communication device in the above method embodiment. The communication device 1200 may be a third communication device or a component that may be configured in a third communication device. The communication unit 1201 is used to perform the sending-related operations on the third communication device side in the above method embodiment, and the processing unit 1202 is used to perform the processing-related operations on the third communication device side in the above method embodiment.

As an example, the communication device 1200 is used to execute the actions executed by the third communication device in the embodiment shown in FIG. 4A above.

For example, the communication unit 1201 is used to receive training capability indication information of a first communication device; the training capability indication information is used to indicate that the first communication device supports multi-model training; and receive reasoning requirement information and reasoning capability information of a second communication device; the reasoning capability information includes reasoning capability indication information, and the reasoning capability indication information is used to indicate that the second communication device supports multi-model reasoning;

The communication unit 1201 is further configured to send model request information to the first communication device, wherein the model request information includes the reasoning requirement information; and receive first information from the first communication device, wherein the first information includes information of a first model, wherein the first model is a multi-model, and the first model is determined according to the reasoning requirement information;

The communication unit 1201 is further configured to send second information to the second communication device, where the second information includes information of the first model.

Based on the same technical concept, the embodiment of the present application also provides a communication device, as shown in FIG13, which is a schematic diagram of a communication device provided by the present application. The communication device 1300 can be the first communication device, the processor of the first communication device, or the chip in the above embodiment. The communication device 1300 can be used to perform the operation performed by the first communication device in the above method embodiment. The communication device 1300 includes: a processor 1302. Optionally, the communication device 1300 can also include a communication interface 1301, a memory 1303, and a communication bus 1304. Among them, the communication interface 1301, the processor 1302, and the memory 1303 can be connected to each other through the communication bus 1304; the communication bus 1304 can be a peripheral component interconnect standard (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The communication bus 1304 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, FIG13 shows only one thick line, but this does not mean that there is only one bus or one type of bus.

Processor 1302 may be a CPU, a microprocessor, an ASIC, or one or more integrated circuits for controlling the execution of the program of the present application.

The communication interface 1301 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), wired access networks, etc.

The memory 1303 may be a ROM or other type of static storage device that can store static information and instructions, a RAM or other type of dynamic storage device that can store information and instructions, or an electrically erasable programmable read-only memory (EPROM). Programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by the computer, but is not limited to this. The memory can be independent and connected to the processor through the communication bus 1304. The memory can also be integrated with the processor.

The memory 1303 is used to store computer-executable instructions for executing the solution of the present application, and the execution is controlled by the processor 1302. The processor 1302 is used to execute the computer-executable instructions stored in the memory 1303, thereby realizing the communication method provided in the above embodiment of the present application.

Optionally, the computer-executable instructions in the embodiments of the present application may also be referred to as application code, which is not specifically limited in the embodiments of the present application.

FIG14 is a schematic diagram of the device structure of a chip provided in an embodiment of the present application. The chip 1400 includes an interface circuit 1401 and one or more processors 1402. Optionally, the chip 1400 may also include a bus. Wherein: the processor 1402 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above-mentioned eye tracking method can be completed by an integrated logic circuit of hardware in the processor 1402 or instructions in the form of software. The above-mentioned processor 1402 may be a general-purpose processor, a digital communicator (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The various methods and steps disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

The interface circuit 1401 can be used to send or receive data, instructions or information. The processor 1402 can use the data, instructions or other information received by the interface circuit 1401 to process, and can send the processing completion information through the interface circuit 1401.

Optionally, the chip further includes a memory 1403, which may include a read-only memory and a random access memory, and provides operation instructions and data to the processor. A portion of the memory 1403 may also include a non-volatile random access memory (NVRAM).

Optionally, the memory stores executable software modules or data structures, and the processor can perform corresponding operations by calling operation instructions stored in the memory (the operation instructions can be stored in the operating system).

Optionally, the chip can be used in the first communication device (second communication device, third communication device) involved in the embodiment of the present application. Optionally, the interface circuit 1401 can be used to output the execution result of the processor 1402. The communication method provided by one or more embodiments of the present application can refer to the aforementioned embodiments, which will not be repeated here.

It should be noted that the corresponding functions of the interface circuit 1401 and the processor 1402 can be implemented through hardware design, software design, or a combination of hardware and software, and there is no limitation here.

An embodiment of the present application also provides a computer-readable storage medium, on which computer instructions for implementing the method executed by the first communication device in the above method embodiment are stored, and/or computer instructions for implementing the method executed by the second communication device in the above method embodiment are stored, and/or computer instructions for implementing the method executed by the third communication device in the above method embodiment are stored.

For example, when the computer program is executed by a computer, the computer can implement the method performed by the first communication device in the above method embodiment.

An embodiment of the present application also provides a computer program product comprising instructions, which, when executed by a computer, enables the computer to implement the method performed by the first communication device in the above method embodiment, and/or when executed by a computer, enables the computer to implement the method performed by the second communication device in the above method embodiment, and/or when executed by a computer, enables the computer to implement the method performed by the third communication device in the above method embodiment.

An embodiment of the present application also provides a chip device, including a processor, for calling a computer program or computer instruction stored in the memory so that the processor executes a communication method of the embodiment shown in FIG. 4A or FIG. 4B above.

In a possible implementation, the input of the chip device corresponds to the receiving operation in the embodiment shown in FIG. 4A or FIG. 4B , and the output of the chip device corresponds to the sending operation in the embodiment shown in FIG. 4A or FIG. 4B .

Optionally, the processor is coupled to the memory via an interface.

Optionally, the chip device further comprises a memory, in which computer programs or computer instructions are stored.

The processor mentioned in any of the above places may be a general-purpose central processing unit, a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of a program of a communication method of the embodiment shown in FIG. 4A or FIG. 4B. The memory mentioned in any of the above places may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM), etc.

It should be noted that, for the sake of convenience and brevity of description, the explanation of the relevant contents and beneficial effects in any of the communication devices provided above can refer to the corresponding eye tracking method embodiments provided above, and will not be repeated here.

In the present application, the communication devices may also include a hardware layer, an operating system layer running on the hardware layer, and an application layer running on the operating system layer. Among them, the hardware layer may include hardware such as a central processing unit (CPU), a memory management unit (MMU), and a memory (also called main memory). The operating system of the operating system layer may be any one or more computer operating systems that implement business processing through processes, such as Linux operating system, Unix operating system, Android operating system, iOS operating system, or Windows operating system. The application layer may include applications such as browsers, address books, word processing software, and instant messaging software.

The division of modules in the embodiments of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional module in each embodiment of the present application may be integrated into a processor, or may exist physically separately, or two or more modules may be integrated into one module. The above-mentioned integrated modules may be implemented in the form of hardware or in the form of software functional modules.

Through the description of the above implementation mode, it can be clearly understood by those skilled in the art that the embodiments of the present application can be implemented by hardware, firmware, or a combination thereof. When software is used for implementation, the above functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein the communication media include any medium that facilitates the transmission of a computer program from one place to another. The storage medium can be any available medium that a computer can access. Taking this as an example but not limited to: a computer-readable medium may include RAM, ROM, electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer. In addition. Any connection can be appropriately a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, wireless, and microwave are included in the fixation of the medium. As used in the embodiments of the present application, disk and disc include compact disc (CD), laser disc, optical disc, digital video disc (DVD), floppy disk, and Blu-ray disc, where disks usually copy data magnetically and discs use lasers to copy data optically. The above combinations should also be included in the scope of protection of computer-readable media.

In short, the above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made according to the disclosure of the present application shall be included in the protection scope of the present application.

Claims

A communication method, comprising:

The first communication device receives model request information, wherein the model request information includes reasoning requirement information;

The first communication device determines a first model according to the model request information, wherein the first model is a multi-model;

The first communication device sends first information, where the first information includes information of the first model.
The method according to claim 1, characterized in that before the first communication device receives the model request information, the method further comprises:

The first communication device sends training capability indication information, where the training capability indication information is used to indicate that the first communication device supports multi-model training.
The method according to claim 1 is characterized in that when the model request information is used to request training of multiple models, the model request information also includes a training strategy for the multiple models; the first communication device determines the first model according to the model request information, comprising: the first communication device performs training according to the inference requirement information and the training strategy for the multiple models to obtain multiple sub-models of the first model; or

When the model request information is used to request training of multiple models; the first communication device determines the first model according to the model request information, including: the first communication device determines the training strategy of the multiple models according to the reasoning requirement information; and performs training according to the reasoning requirement information and the training strategy of the multiple models to obtain multiple sub-models of the first model;

The multi-model training strategy includes one or more of the following:

Data processing strategy, training algorithm, training mode, number of sub-models, and type of sub-models.
The method according to claim 1, wherein when the model request information is used to request to obtain multiple models, the first communication device determines the first model according to the model request information, comprising:

The first communication device determines the first model from at least one preset multiple models according to the inference requirement information.
The method according to any one of claims 1 to 4 is characterized in that the model request information also includes multi-model indication information, and the multi-model indication information is used to indicate that the model requested to be trained or obtained is a multi-model.
The method according to any one of claims 1 to 5, characterized in that the reasoning requirement information includes one or more of the following:

The type of inference, the performance requirements of inference, the speed requirements of inference, and the power consumption requirements of inference.
The method according to any one of claims 1 to 6, characterized in that the information of the first model includes model information of the first model and information of multiple sub-models of the first model, and the information of each sub-model includes one or more of the following:

Sub-model identification information, sub-model level, sub-model performance, and performance constraints;

The multiple sub-models include multiple first-level sub-models and one second-level sub-model, and the second-level sub-model is used to aggregate the reasoning information of the multiple first-level sub-models; or the multiple sub-models are all first-level sub-models, and the information of the first model also includes aggregation method and/or weight information.
The method according to any one of claims 1 to 7, characterized in that the method further comprises:

The first communication device sends the reasoning performance information of the first model, where the reasoning performance information of the first model includes one or more of the following:

The performance of the first model, the size information of the first model, the power consumption of the reasoning of the first model, the reasoning speed of the first model, and the computing power of the first model.
The method according to any one of claims 1 to 8, characterized in that the first communication device is any one of the following:

Model training function network element, model training function entity, and communication device including model training function.
A communication method, characterized by comprising:

The second communication device receives second information, wherein the second information includes information of a first model, the first model is determined according to reasoning requirement information, and the first model is a multi-model;

The second communication device obtains inference information of the first model based on the information of the first model.
The method according to claim 10, characterized in that before the second communication device receives the second information, the method further comprises:

The second communication device sends the reasoning capability information and the reasoning requirement information;

The reasoning capability information includes reasoning capability indication information and one or more of the following:

The computing power and storage space for reasoning; the reasoning capability indication information is used to indicate that the second communication device supports multi-model reasoning;

The reasoning requirement information includes one or more of the following:

The type of inference, the performance requirements of inference, the speed requirements of inference, and the power consumption requirements of inference.
The method according to claim 10, characterized in that the information of the first model includes model information of the first model and information of multiple sub-models of the first model, and the information of each sub-model includes one or more of the following:

Sub-model identification information, sub-model level, sub-model performance, and performance constraints.
The method according to claim 12, characterized in that the multiple sub-models include multiple first-level sub-models and one second-level sub-model, and the second-level sub-model is used to aggregate the reasoning information of the multiple first-level sub-models;

The second communication device obtains the inference information of the first model based on the information of the first model, including:

The second communication device performs reasoning using the multiple first-level sub-models respectively based on the information of the multiple first-level sub-models to obtain reasoning information of the multiple first-level sub-models;

The second communication device aggregates the reasoning information of the plurality of first-level sub-models using the second-level sub-model to obtain the reasoning information of the first model.
The method according to claim 12, characterized in that the multiple sub-models are all first-level sub-models, and the information of the first model also includes aggregation method and/or weight information;

The second communication device obtains the inference information of the first model based on the information of the first model, including:

The second communication device performs reasoning using the multiple sub-models respectively based on the information of the multiple sub-models to obtain reasoning information of the multiple sub-models;

The second communication device aggregates the reasoning information of the multiple sub-models according to the aggregation method and/or weight information to obtain the reasoning information of the first model.
The method according to any one of claims 10 to 14, characterized in that the second communication device is any one of the following:

Model reasoning function network element, model reasoning function entity, and communication device including model reasoning function.
A communication method, comprising:

The third communication device receives the training capability indication information of the first communication device; the training capability indication information is used to indicate that the first communication device supports multi-model training;

The third communication device receives the reasoning requirement information and reasoning capability information of the second communication device; the reasoning capability information includes reasoning capability indication information, and the reasoning capability indication information is used to indicate that the second communication device supports multi-model reasoning;

The third communication device sends model request information to the first communication device, where the model request information includes the reasoning requirement information;

The third communication device receives first information from the first communication device, where the first information includes information of a first model, the first model is a multi-model, and the first model is determined according to the reasoning requirement information;

The third communication device sends second information to the second communication device, where the second information includes information of the first model.
The method according to claim 16, characterized in that the inference requirement information of the second communication device includes one or more of the following:

The type of inference, the performance requirements of inference, the speed requirements of inference, and the power consumption requirements of inference.
The method according to claim 16, characterized in that the reasoning capability information of the second communication device further includes one or more of the following:

The computing power and storage space for reasoning.
The method according to claim 16 is characterized in that the model request information also includes multi-model indication information, and the multi-model indication information is used to indicate that the model requested to be trained or obtained is a multi-model.
The method according to claim 16, characterized in that the method further comprises:

The third communication device receives the inference performance information of the first model from the first communication device;

The third communication device adjusts the number of sub-models in the first model according to the reasoning requirement information and the reasoning capability information of the second communication device, the reasoning information of the first model and the information of the first model;

The inference performance information of the first model includes one or more of the following:

The performance of the first model, the size information of the first model, the power consumption of the reasoning of the first model, the reasoning speed of the first model, and the computing power of the first model.
The method according to any one of claims 16 to 20, characterized in that the information of the first model includes model information of the first model and information of multiple sub-models of the first model, and the information of each sub-model includes one or more of the following:

Sub-model identification information, sub-model level, sub-model performance, and performance constraints;

The multiple sub-models include multiple first-level sub-models and one second-level sub-model, and the second-level sub-model is used to aggregate the reasoning information of the multiple first-level sub-models; or the multiple sub-models are all first-level sub-models, and the information of the first model also includes aggregation method and/or weight information.
The method according to any one of claims 16 to 21, characterized in that the third communication device is any one of the following:

Model management function network element, model management function entity, and communication device including model management function.
A communication device, characterized in that it includes a unit or module for executing the method as described in any one of claims 1 to 9, or a unit or module for executing the method as described in any one of claims 10 to 15, or a unit or module for executing the method as described in any one of claims 16 to 22.
A communication device, characterized in that it includes a processor and an interface circuit, wherein the interface circuit is used to receive signals from other communication devices outside the communication device and transmit them to the processor or send signals from the processor to other communication devices outside the communication device, and the processor is used to implement the method as described in any one of claims 1 to 9, or to implement the method as described in any one of claims 10 to 15, or to implement a unit or module of the method as described in any one of claims 16 to 22 through a logic circuit or execution code instructions.
A computer program product, characterized in that it includes a computer program, and when the computer program is executed by a communication device, it implements the method as described in any one of claims 1 to 9, or implements the method as described in any one of claims 10 to 15, or implements the method as described in any one of claims 16 to 22.
A computer-readable storage medium, characterized in that a computer-readable program or instruction is stored in the storage medium, and when the computer program or instruction is executed by a communication device, the method as described in any one of claims 1 to 9 is implemented, or the method as described in any one of claims 10 to 15 is implemented, or the method as described in any one of claims 16 to 22 is implemented.