O-RAN WG2 AI/ML工作流程与标准要求（v01.03.02）

需积分: 10 142 浏览量更新于2024-07-08 收藏 2.08MB DOCX 举报

本文档是O-RAN Working Group 2（WG2）关于人工智能/机器学习（AI/ML）在O-RAN（开放Radio接入网）系统中的工作流程描述和要求的技术报告。O-RAN是一个旨在推动开放无线网络架构标准化的联盟，其目标是通过共享硬件和软件组件来提高网络的灵活性和效率。该技术报告v01.03.02由O-RAN Alliance于2021年发布，所有使用、访问或下载此规范文档的行为均需遵循O-RAN Adopter License Agreement的条款，可在附件ZZZ中找到。第一章"Introduction"（介绍）部分定义了报告的范围，概述了AI/ML在O-RAN中的应用背景，以及引用的相关标准和文献。同时，还提供了必要的定义和缩写列表，如AI/ML的具体含义及其在O-RAN中的术语使用。第二章"Machine Learning"（机器学习）详细阐述了通用术语和定义，包括AI的不同类型，强调了如何在无线通信领域应用这些技术。这一章节介绍了AI的三个主要类别：监督学习、无监督学习和强化学习，这些都是O-RAN架构中可能采用的基础算法。第三章"Types of Machine Learning algorithms"（机器学习算法类型）深入探讨了每种类型的机器学习方法在O-RAN中的具体应用，如如何利用监督学习进行数据预测和分类，无监督学习进行模式识别和聚类，以及强化学习如何通过试错学习优化网络决策。第四章"Procedure/Integration"（流程/整合）可能涵盖了AI/ML如何与O-RAN的控制循环（control loops）相结合，包括AI功能在无线网络中的部署策略、数据处理流程以及如何确保性能监控和可解释性。这部分内容可能会涉及AI在无线资源管理、网络优化、故障检测和自我修复等方面的具体实施步骤。这份文档深入探讨了AI/ML在O-RAN NR无线通信标准协议中的角色和应用，提供了技术框架和实施指南，对O-RAN的运营商、制造商和研究者来说，是一份理解如何利用AI驱动无线网络现代化的重要参考资料。通过阅读和理解这份报告，参与者可以更好地设计、集成和优化基于AI的解决方案，以提升O-RAN系统的整体性能和效率。

ML (model) Life-cycle: The life-cycle of the ML model

includes deployment, instantiation and termination of ML

model components.

These are operational phases: the initial

training, inference, possible re-training

ML Pipeline: The set of functionalities, functions, or functional

entities specific for an ML-assisted solution.

a ML pipeline may consist of one or

several data sources in a data pipeline, a

model training pipeline, a model

evaluation pipeline and an actor.

ML Training Host: The network function which hosts the

training of the model, including offline and online training

Non-RT RIC can also be a training host.

ML training can be performed offline

using data collected from the RIC, O-

DU and O-RU.

ML Inference Host: The network function which hosts the ML

model during inference mode (which includes both the model

execution as well as any online learning if applicable).

The ML inference host often coincides

with the Actor. The ML-host informs the

actor about the output of the ML

algorithm, and the Actor takes a decision

for an action.

Actor: The entity which hosts an ML assisted solution using

the output of ML model inference.

Action: An action performed by an actor as a result of the

output of an ML assisted solution.

Subject of Action: The entity or function which is configured,

controlled, or informed as result of the action.

Model training information: Information needed for training

the ML model.

This is the data of the ML model

including the input plus optional labels

for supervised training

Model inference information: Information needed as input for

the ML model for inference.

The data needed by an ML model for

training and inference may largely

overlap, however they are logically

different.

Model compiling info: Information needed for compiling the

ML model.

It includes the trained ML model ,

model inference host info, and other

possible requirements for model

compiling, e.g., acceptable accuracy loss

thresholds, any specific operations to be

O-RAN.WG2.AIML-v01.03.02

performed.

Model inference host info: Information of the inference host

needed as model compiling info.

This may include information e.g.,

network bandwidth, memory and

computing capabilities of the inference

host.

Model Compiling Host: Optional network function which

compiles the trained model into specific format for optimized

inference execution in an inference host based on the

compiling information .

Model compiling involves hardware-

specific optimization to achieve

improved computing and memory

efficiency. According to different

deployment methods, the compiled

model could be published to the model

store/ management module as a

containerized image (including

Inference Engine and compiled models)

or a compiled model file format ， and

then deployed to the inference host.

Non-RT RIC can also be a compiling

host. ML compiling can be performed

offline.

Model Inference Engine: The specific inference framework

used to run a compiled model in the inference host.

The functions of model inference engine

includes parsing model files, splitting

operations, and executing inference

instruction stream to finish ML model

inference calculation and return

inference result.

Model Management: The network function that manages the

ML models to be deployed in the inference host.

Model management manages models

that are onboarded directly from ML

training host or those from ML

compiling jost when model comiling is

executed after training.

Data Discovery: Data discovery uses smart tools to collect data

from multiple sources and then consolidate them to a single

source for easy use.

The general process includes data

collecting, data cleansing, data loading,

data transforming, data mining and data

visualization

Data Augmentation: Data augmentation could be used to

enhance the model generalization and reduce overfitting,

especially when dataset is limited or imbalance exists.

There are two types of data:

Time series data: augmentation with

consideration of event distribution

O-RAN.WG2.AIML-v01.03.02

pattern in temporal space

Non-time series data: augmentation with

physical constraint

Data Labelling: Data labelling is a time-consuming task and

needs domain knowledge. The platform should provide or

integrate various labelling/ auto-labelling tools for

offline/online annotation.

The labelled data should be validated

crossly, and some label mistakes should

be eliminated

Feature Engineering: At the beginning of model design, we

need to study and define what kinds of features could be learnt

to represent the objectives or data, and what kinds of actions

should be used to for better prediction performance.

Design the feature extraction

mechanism, such as learning from

multiple modalities, or learning from

single modality, etc. Meanwhile, we

need to define which features could be

leveraged and which could be fused with

other features.

Nowadays, end2end system building is

emerging also, but raw data is also a

kind of feature in this context.

This is an optional operation for model

training.

Model Selection: When initiating a training task, ML designer

will assess the cost and resources for training or inference, such

as hardware platform to inference, GPU memory, inference

speed, training accuracy, training time, etc. Based on those

requirements analysis, the system can select the corresponding

configurations for training or inference.

For model design, the following factors

should be considered:

 ML meta architecture for

specific tasks

 ML/DL framework

(pytorch, tensorflow, caffe,

etc.)

 Data format of input and

output

 Requirements on model

performance (accuracy,

responding time, real-time

factor, etc.)

 Model footprint and HW

platform (ARM, GPU,

CPU, FPGA, etc.)

 Task requirements

Model Optimization: Model optimization refers to the efforts For example, using auto machine

O-RAN.WG2.AIML-v01.03.02

剩余59页未读，继续阅读

AJ_chenrui

粉丝: 4
资源: 7

O-RAN WG2 AI/ML工作流程与标准要求（v01.03.02）

ORAN Spec & Protocol 合集

cn-biz.rar_AIML中文_AIML中文语料库_aiml_cn_aiml语料_aiml语料库

HSG-AIML.github.io:HSG-AIML研究网站

python 3.8安装库 -分享.docx

AIML_Bot-0.0.3-py3-none-any.whl.zip

AIML_Sets-1.0.1-py3-none-any.whl.zip

AIML_Bot-0.0.1-py3-none-any.whl.zip

AIML_Bot-0.0.2-py3-none-any.whl.zip

AIML_Bot_API-0.0-py3-none-any.whl.zip

Python库 | aiml-0.8.6.tar.gz

最新资源