5G/NR - AI/ML  

 

 

 

AI/ML

Finally AI/ML is getting into 3GPP -:). This note is not about what is AI/ML. I have a whole list of section about AI/ML itself and 3GPP also posted a short article about AI/ML. This note is more of how and where AI/ML will be used in 3GPP framework.

3GPP AI/ML Framework

As of now (Dec 2021), it is not clearly decided exactly what kind of AI/ML framework will be used in terms of 5G/NR network. One possible high level view is as shown below (this is based on TR 38.817 V1.0.0 (2021-12) - Figure 4.2-1: Functional Framework for RAN Intelligence).

 

 

Who is doing the data preparation (data pre-processing and cleaning, formatting, and transformation) ?

: Data Collection component just collect the raw data and does not process it for other component to use it. The data processing / preparation is done by the component that comsumes the data. For example, if the collected data is for training model,  Data Collection module just transfer the raw data to Model Training component and the Model Training component preprocess the data. If if the collected data is for model inference,  Data Collection module just transfer the raw data to Model Inference component and the Model Inference component prepare the data.

Principles of 3GPP AI/ML

When I was thinking of applying AI/ML in cellular technoloty (especially on RAN side), I had so many questions poping up in my head. What kind of neural network model will be used ? How it will be trained ? Who (gNB or Corenetwork) will train the network ? What kind of data will be used for training the network ? What kind of use cases will there be ? and so on.. and on ... and on. As of now (Dec 2021), I don't have clear answer to these questions... I would need to wait until the 3GPP specification of AI/ML is finalized (Rel 18). But at least I can get a glimpse of general principles of AI/ML implmentation from TR 38.817 - 4.1. It is stated as follows (Just reading these bullets gives me pretty good idea).

  • The detailed AI/ML algorithms and models for use cases are implementation specific and out of RAN3 scope.
  • The study focuses on AI/ML functionality and corresponding types of inputs/outputs.
  • The input/output and the location of the Model Training and Model Inference function should be studied case by case.
  • The study focuses on the analysis of data needed at the Model Training function from Data Collection, while the aspects of how the Model Training function uses inputs to train a model are out of RAN3 scope.
  • The study focuses on the analysis of data needed at the Model Inference function from Data Collection, while the aspects of how the Model Inference function uses inputs to derive outputs are out of RAN3 scope.
  • Where AI/ML functionality resides within the current RAN architecture, depends on deployment and on the specific use cases.
  • The Model Training and Model Inference functions should be able to request, if needed, specific information to be used to train or execute the AI/ML algorithm and to avoid reception of unnecessary information. The nature of such information depends on the use case and on the AI/ML algorithm.   
  • The Model Inference function should signal the outputs of the model only to nodes that have explicitly requested them (e.g. via subscription), or nodes that are subject to actions based on the output from Model Inference.
  • An AI/ML model used in a Model Inference function has to be initially trained, validated and tested before deployment.
  • The generalized workflow should not prevent to “think beyond” the workflow if the use case requires so.
  • User data privacy and anonymisation should be respected during AI/ML operation.

In short, these can be summarized as

  • 3GPP would not define exact neural network definition / implementation. It would define only the interface and type of input/output of the neural network(Model).
  • The whole AI/ML functionality would be comprised of several different compoments (e.g, Data Collection, Model Training, Model Interence, Actor). Where these components will be located will vary depending on specific use cases.

Workflow of 5G AL/ML

 

< 5G AI/ML workflow based on TR 28.908 V1.2.0 - 4.3 AI/ML workflow for 5GS >

What kind of data to be collected for training ?

As most of you already understand, in order to train a AI/ML system. You need to have clear answer to followings :

  • What kind of data will be used ?
  • How the data would be aquired ?
  • In what type of format the data should be prepared ?

Answers to the first two questions are described in TR 28.908 (v1.2.0) - 5.1.1 Event data for ML training and I think the answer to the third question will be described in various TS.  In short, according to TR 28.908, the data is called 'Event data' implying that those data are to indicates a certain type of events and collected by specific event handler.

 

Followings are some of the example of these events

 

< TR28.908 (v1.2.0) - Table 5.1.1.4-1: Examples of potential Metric Threshold Crossing events >

Event-source Function

Event Name

Event description

Input/Metric

MOI

Condition

Threshold

Monitoring Period

BTSs; NBs; eNBs; gNBs; BSC; RNC

High Call Drop Rate event

Call Drop Rate more than a configurable threshold

Call Drop Rate

Cell

greater than

2 %

15 minutes

BTSs; NBs; eNBs; gNBs; BSC; RNC

Low Availability KPIs event

Availability KPIs dropping below a configurable threshold

Availability

Cell

less than

99 %

30 minutes

BTSs; NBs; eNBs; gNBs; BSC; RNC

Low Retainability KPIs event

Retainability KPIs dropping below a configurable threshold

Retainability

Cell

less than

98 %

30 minutes

BTSs; NBs; eNBs; gNBs; BSC; RNC

High Traffic event

Traffic greater than a configurable threshold

Cell load / PRB Utilization

Cell

greater than

80 %

15 minutes

BTSs; NBs; eNBs; gNBs; BSC; RNC

Interference event

User has experienced interference

SINR

Cell

less than

X dB

15 minutes

BTSs; NBs; eNBs; gNBs; BSC; RNC

Serving cell

-

RSRP

Cell

greater than

Y dBm

15 minutes

Event - source Function : Network Components that collects the data

MOI :  Managed Object Instance. A part of a network that can be managed independently. This could be a physical component, such as a cell in a cellular network, or a logical component, such as a service or a process.

 

 

< TR28.908 (v1.2.0) - Table 5.1.1.4-2: Examples of potential Object-Status Change events >

Event-source Function

Event Name

Event description

MOI

Affected unit / parameter

Change / value

BTSs; NBs; eNBs; gNBs; BSC; RNC; NMS

HW Upgrade event

System Module HW version upgraded

BTS

System Module, Radio Module, …

HW version

SW Upgrade event

System Software Upgraded

BTS

System Module, Radio Module, …

SW version

Capability Enablement event

A specific Capability Enabled on the MOI

BTS

Spectrum Sharing

Spectrum Range for affected RATs

New Sector Addition Event

A new Sector getting added to a Site

Cell

Capacity and Coverage

Number of Sectors / Cells

Network Management

Parameter change event

CM Parameter changes applied for specific network element

Cell

Configuration Parameter

Parameter value

Home status event

MOI (e.g. site) Re-homing

Site

BSC/RNC, OSS

New BSC/ RNC/ OSS

SON, Analytics function

New Site event

New Site Integrated

Site/ gNB

C-SON Functions

Optimization Parameters in C-SON

Predicted Congestion

Trigger for Load Balancing detected

Cell

C-SON LBO

Mobility Parameter changes

Frequent Handover Failures

Trigger for MRO detected

Cell

C-SON MRO

Handover Parameter Changes

PCI Conflict

PCI conflict detected

Cell

C-SON PCI

PCI Changes

PRACH Conflict

PRACH conflict detected

Cell

C-SON PRACH

PRACH related parameter Changes

NCR Change

New First tier neighbour getting added

Cell

C-SON ANR

NCR Changes

Frequency Layer Change

New Frequency Layer added onto a site

BTS

C-SON

Frequency Layer Addition

What is it used for in 5G ?

Following use cases are based on TR 38.817 (V17.0.0 (2022-04)). As you would notice, it seems that 3GPP is targetting mostly for Network side optimization for now.

Network Energy Saving

The informations to be collected are as follows :

  • From local node:
    • UE mobility/trajectory prediction
    • Current/Predicted Energy efficiency
    • Current/Predicted resource status
  • From the UE:
    • UE location information (e.g., coordinates, serving cell ID, moving velocity) interpreted by gNB implementation when available
    • UE measurement report (e.g., UE RSRP, RSRQ, SINR measurement, etc), including cell level and beam level UE measurements
  • From neighbouring NG-RAN nodes:
    • Current/Predicted energy efficiency
    • Current/Predicted resource status
    • Current energy state (e.g., active, high, low, inactive)

Load Balancing

The informations to be collected are as follows :

  • From the local node:
    • Current and predicted own resource status
    • UE trajectory prediction
    • Current and predicted UE traffic
    • Predicted resource status information of neighbouring NG-RAN node(s)
  • From the UE:
    • UE location information (e.g., coordinates, serving cell ID, moving velocity) interpreted by gNB implementation when available
    • UE Mobility History Information
    • UE measurement report (e.g., UE RSRP, RSRQ, SINR measurement, etc), including cell level and beam level UE measurements
  • From neighbouring NG-RAN Nodes:
    • Current and predicted resource status
    • UE performance measurement at traffic offloaded neighbouring cell

Mobility Optimization

The informations to be collected are as follows :

  • From the UE:
    • UE location information (e.g., coordinates, serving cell ID, moving velocity) interpreted by gNB implementation when available.
    • Radio measurements related to serving cell and neighbouring cells associated with UE location information, e.g., RSRP, RSRQ, SINR.
    • UE Mobility History Information.
  • From the neighbouring RAN nodes:
    • UE’s history information from neighbour
    • Position, QoS parameters and the performance information of historical HO-ed UE (e.g., loss rate, delay, etc.)
    • Current/predicted resource status
    • UE handovers in the past that were successful and unsuccessful, including too-early, too-late, or handover to wrong (sub-optimal) cell, based on existing SON/RLF report mechanism.
  • From the local node:
    • UE trajectory prediction
    • Current/predicted resource status
    • Current/predicted UE traffic

Beam Management

The Release 18 RAN1 study on AI/ML for NR Air Interface explores the benefits of augmenting the air interface with features enabling improved support of AI/ML based algorithms for enhanced performance and/or reduced complexity or overhead.

Following is based on R1-2204060. Refer to the original document for further details.

  • CSI feedback enhancement : Enhancements to CSI, such as frequency domain compression, have been agreed upon, with other enhancements like time-domain prediction still under consideration (e.g., overhead reduction, improved accuracy, prediction)
  • Beam management : Use AI and ML techniques to predict the best beam to use for a given user device at a given time and place, accounting for various factors such as user location, movement, and network conditions (i.e, beam prediction in time, and/or spatial domain for overhead and latency reduction, beam selection accuracy improvement)
  • Positioning accuracy enhancements for different scenarios including : Compared to traditional positioning algorithms, machine learning-assisted positioning can require less additional feedback overhead, and positioning estimates can be made directly inside the radio access network (RAN) and handle the situation with heavy NLOS conditions .

NOTE : I am writing a few separte notes about AI/ML on PHY and Air Interface.

Implementation Examples

Even before 3GPP specification process is done, some modem chipset starts coming out equipped with AI functionalities in radio stack.

Qualcomm X65

  • AI-Enhanced Signal Boost (Adaptive antenna tuning solution)

Qualcomm X70

  • AI-based channel-state feedback and dynamic optimization
  • AI-based mmWave beam management for superior mobility and coverage robustness
  • AI-based network selection for superior mobility and link robustness
  • AI-based adaptive antenna tuning for up to 30% improved context detection for higher average speeds and coverage

Qualcomm X75

SamSung

Reference

YouTube

YouTube (Korean)