You are on page 1of 19

MACHINE LEaRNING &

DEEP LEARNING : USE CASES


April 2018
Oscar Real
Contents
1 INTRODUCTION

HIGH RSSI DETECTION: : NEURAL NETWORK


2 (BINARY CLASSIFICATION)
POOR CQI DETECTION : NEURAL NETWORK
3 (MULTI-CLASS CLASSIFICATION)
CELL CONGESTION TRIGGER : RECURRENT
4 NEURAL NETWORK

5 ELASTIC RAN MODEL

6 SUMMARY
Ericsson Internal | April 2018 | Page 2
INTRODUCTION
› The aim is to show you some examples of interesting Machine Learning and Deep Learning
Use Cases giving data from a LTE live network

› These Use Cases were done by myself based on my experience.

› Examples are based RAN data

› Scope is to show some ideas to strengthen ML and DL applications at Radio daily tasks on
our customer projects.

Ericsson Internal | April 2018 | Page 3


Contents
1 INTRODUCTION

HIGH RSSI DETECTION: : NEURAL NETWORK


2 (BINARY CLASSIFICATION)
POOR CQI DETECTION : NEURAL NETWORK
3 (MULTI-CLASS CLASSIFICATION)
CELL CONGESTION TRIGGER : RECURRENT
4 NEURAL NETWORK

5 ELASTIC RAN MODEL

6 SUMMARY
Ericsson Internal | April 2018 | Page 4
NEURAL NETWORK (BINARY CLASS)
USE CASE : HIGH RSSI DETECTION
› The Deep Learning methodology to build the UL PRB USAGE
model in this case for HIGH RSSI detection
1  UL HIGH RSSI
follows the next steps:
0  NO ISSUE
1. Initialize parameters pmRadioRecInterferencePwrPrb01

2. Loop for interactions : pmRadioRecInterferencePwrPrbN

2.1 Forward propagation. BINARY CROSS-ENTROPY COST FUNCTION :


2.2 Compute cost function
2.3 Backward propagation
2.4 Update parameters
Accuracy about predict
3. Use trained parameters to predict labels. Cost Function
labels > 99.8%
Trend constant

Ericsson Internal | April 2018 | Page 5


NEURAL NETWORK (BINARY CLASS)
USE CASE : HIGH RSSI DETECTION
INPUT LAYER OUTPUT LAYER
-95 dBm

-100 dBm

-105 dBm

-110 dBm

-115 dBm

UL PRB USAGE HIGH RSSI

-95 dBm

-100 dBm

-105 dBm

-110 dBm

-115 dBm

UL PRB USAGE
Ericsson Internal | April 2018 | Page 6 HIGH RSSI
Contents
1 INTRODUCTION

HIGH RSSI DETECTION: : NEURAL NETWORK


2 (BINARY CLASSIFICATION)
POOR CQI DETECTION : NEURAL NETWORK
3 (MULTI-CLASS CLASSIFICATION)
CELL CONGESTION TRIGGER : RECURRENT
4 NEURAL NETWORK

5 ELASTIC RAN MODEL

6 SUMMARY
Ericsson Internal | April 2018 | Page 7
NEURAL NETWORK (MULTI-CLASS)
USE CASE : POOR CQI DETECTION
› The Deep Learning methodology to build the
0  ACCEPTABLE
model in this case for Poor CQI detection pmRadioUeRepCqiDistr_01
AVERAGE CQI
follows the next steps: pmRadioUeRepCqiDistr_02
1  LOW AVERAGE CQI
…..
1. Initialize parameters 2  ACCEPTABLE AVERAGE
pmRadioUeRepCqiDistr_15
CQI BUT HIGH
2. Loop for interactions : INTERFERENCE CLOSE

2.1 Forward propagation. TO CELL

2.2 Compute cost function CATEGORICAL CROSS-ENTROPY COST FUNCTION :

2.3 Backward propagation


2.4 Update parameters
3. Use trained parameters to predict labels. Cost Function
Trend constant

Accuracy about predict


labels > 99.9%

Ericsson Internal | April 2018 | Page 8


NEURAL NETWORK (MULTI-CLASS)
USE CASE : POOR CQI DETECTION
0  ACCEPTABLE
AVERAGE CQI
1  LOW AVERAGE CQI

2  ACCEPTABLE AVERAGE
CQI BUT HIGH
INTERFERENCE CLOSE

TO CELL

Input CQI
Output Y_Prediction

Ericsson Internal | April 2018 | Page 9


NEURAL NETWORK (MULTI-CLASS)
USE CASE : POOR CQI DETECTION
0  ACCEPTABLE
AVERAGE CQI
1  LOW AVERAGE CQI

2  ACCEPTABLE AVERAGE
CQI BUT HIGH
INTERFERENCE CLOSE

TO CELL

Input CQI
Output Y_Prediction

Ericsson Internal | April 2018 | Page 10


Contents
1 INTRODUCTION

HIGH RSSI DETECTION: : NEURAL NETWORK


2 (BINARY CLASSIFICATION)
POOR CQI DETECTION : NEURAL NETWORK
3 (MULTI-CLASS CLASSIFICATION)
CELL CONGESTION TRIGGER : RECURRENT
4 NEURAL NETWORK

5 ELASTIC RAN MODEL

6 SUMMARY
Ericsson Internal | April 2018 | Page 11
RECURRENT NEURAL NETWORK
USE CASE : CELL CONGESTION TRIGGER
0  NO CONGESTION
› The Deep Learning methodology to build the
1  ENTERING CONGESTION
model in this case for Poor CQI detection
2  CONGESTION
follows the next steps:
3  LEAVING CONGESTION
1. Initialize parameters
2. Loop for interactions :
2.1 Forward propagation. …..

2.2 Compute cost function


2.3 Backward propagation PRB DL Usage PRB DL Usage

CCE_Utilization CCE_Utilization
2.4 Update parameters …. Avg DL User Throughput
Avg DL User Throughput
Cell DL Throughput
3. Use trained parameters to predict labels. Cell DL Throughput

CATEGORICAL CROSS-ENTROPY COST FUNCTION :

Accuracy about predict labels > 98.2%

Ericsson Internal | April 2018 | Page 12


RECURRENT NEURAL NETWORK
USE CASE : CELL CONGESTION TRIGGER

Input

Output :
0  NO CONGESTION
1  ENTERING CONGESTION
2  CONGESTION
3  LEAVING CONGESTION

Ericsson Internal | April 2018 | Page 13


Contents
1 INTRODUCTION

HIGH RSSI DETECTION: : NEURAL NETWORK


2 (BINARY CLASSIFICATION)
POOR CQI DETECTION : NEURAL NETWORK
3 (MULTI-CLASS CLASSIFICATION)
CELL CONGESTION TRIGGER : RECURRENT
4 NEURAL NETWORK

5 ELASTIC RAN MODEL

6 SUMMARY
Ericsson Internal | April 2018 | Page 14
ELASTIC RAN APPROACH
List of Cells with other Cells weights
• Downlink CARRIER AGGREGATION : for UL_COMP and DL CA

For a given Cell_i, its weight for a Cell_j in different List of Cells List of Cells
LTE frequency is: with their external Cells with their internal Cells
𝑆𝑎𝑚𝑝𝑙𝑒𝑠 ¿
𝑊𝑒𝑖𝑔h 𝑡 𝐶𝑒𝑙 𝑙 = ∙ 𝐴𝑣𝑔 ¿
𝑗 𝐾 =𝑁
List of Cells with their external Cells
∑ 𝑆𝑎𝑚𝑝𝑙𝑒𝑠 ¿
𝑘=1
Order by their weights

• UL COMP : List of Cells with their external Cells


Order by their weights while the sum of all eNB <7
For a given Cell_i, its weight for a Cell_j in the same Or External Cells <109 for a given eNB

LTE frequency is:


𝑆𝑎𝑚𝑝𝑙𝑒𝑠 ¿
𝑊𝑒𝑖𝑔h 𝑡 𝐶𝑒𝑙 𝑙 = 𝐾 =𝑁
∙ 𝐴𝑣𝑔 ¿
𝑗
Check if eNode_A
∑ 𝑆𝑎𝑚𝑝𝑙𝑒𝑠 ¿ Join both tables :
𝑘= 1
and eNode_B has Yes
External and internal
Bidirectional
where LSA is Propagation Loss : Cell Relations
Assignment

LSAcell_j
No
Ericsson Internal | April 2018 | Page 15
Remove ESCELL_Relation
ELASTIC RAN APPROACH
› A Elastic-RAN study has been done in an operator in Sao Paulo Area. Next, it is shown the E5 connections
between different eNBs :

SAO PAULO AREA SAO PAULO METROPOLITAN AREA

Ericsson Internal | April 2018 | Page 16


Contents
1 INTRODUCTION

HIGH RSSI DETECTION: : NEURAL NETWORK


2 (BINARY CLASSIFICATION)
POOR CQI DETECTION : NEURAL NETWORK
3 (MULTI-CLASS CLASSIFICATION)
CELL CONGESTION TRIGGER : RECURRENT
4 NEURAL NETWORK

5 ELASTIC RAN MODEL

6 SUMMARY
Ericsson Internal | April 2018 | Page 17
Summary
› SOME USE CASES HAVE BEEN PRESENTED :
› UPLINK HIGH RSSI
› POOR DL QCI DETECTION
› CELL CONGESTION TRIGGER
› ELASTIC RAN MODEL

› FUTURE ACTIONS : TRY TO TURN DEEP-LEARNING INTO Q-DEEP-


LEARNING (OR ARTIFICIAL INTELLIGENCE).
› DECTECT AND TAKE ACTIONS THE SYSTEM ITSELF TO
IMPROVE END-USER PERFORMANCE WITHOUT A HUMAN
INTERVENTION.

› MACHINE LEARNING WILL IMPROVE OUR PROCESSES, DELIVERY


QUALITY AND BRING EFFICIENCIES.

Ericsson Internal | April 2018 | Page 18


Ericsson Internal | April 2018 | Page 19

You might also like