Machine Intelligence for Telecom and Beyond

Transcription

1 Machine Intelligence for Telecom and Beyond Elena Fersman, PhD Research Director, Machine Intelligence and Automation, Ericsson Adjunct Professor, Cyber-Physical Systems, KTH

2 Ericsson Mobile Infrastructure, Digital Services, Managed Services 180 countries 100K employees patents 1 bn subscribers on networks managed by us 5G 50% are PhDs Ericsson Research 2G, 3G, 4G and 5G were invented at Ericsson Research 40% of all Ericsson patents come from Ericsson Research Elena Fersman Head of Research Area Machine Intelligence and Automation, Ericsson Research Adjunct Professor in Cyber-Physical Systems specialized in Automation, KTH fersman.blogspot.com, instagram: elenafersman

3 Do you know what s common between these? Cortana Lucida Mika Amelia Alexa Siri Watson

4 Data Lake Cortana Lucida Mika Amelia Alexa Siri Watson

5 Hi, what is Machine Intelligence? Hi Elena, according to Ericsson definition, Machine Intelligence is a combination of MachineLearning and Artificial Intelligence.

6 What s MSDP? MSDP = Multicast Source Discovery Protocol MSDP = Multiservice Delivery Platform MSDP = Managed Services Delivery Platform

7 What are your sources? Ericsson Product Catalogue, CPI Store, Wikipedia, and your calendar

8 Tell me when site X is likely to require a preventative maintenance visit? Judging from the site profile learned over time, and the current site behavior, the maintenance will be needed in Q

9

10 What actions did you do for Operator A s network last month? I changed parameter configuration in the 5G network nodes. I also sent offers to subscribers with high risk of churn.

11 Why? Because we need to stay competitive towards our enterprise customers while keeping our subscribers happy.

12 The value is in data Cortana Lucida Mika Amelia Alexa Siri Watson

13 From raw data to action

14 Data and its processing

15 What s in the data lake?

16 ONE DAY IN THE LIFE OF A MEDIUM SIZED NETWORK (~10M CUSTOMERS) Web pages 700,000,000 Internet sessions (PDP) 66,000,000 Videos 40,000,000 Radio sessions (RAB) 120,000, more types of events Handovers (HSDSCH-CC) 300,000,000 Sum data 10à100 TB/day Real-time data rate 100,000à1,000,000 events/second

17 5G a foundation for digitalization 1G 2G 3G 4G 5G cutting the cord, adding mobility going digital adding video & data mobile broadband 29 billion devices Telephony and mobile broadband billion people 1 billion places A digital infrastructure for industrial and societal transformation

18 SENSORS EVERYWHERE 5g USE CASES BROADBAND AND MEDIA EVERYWHERE SMART VEHICLES, TRANSPORT INFRASTRUCTURE, MONITOR AND CONTROL CRITICAL CONTROL OF REMOTE DEVICES INTERACTION HUMAN-IOT

19 Evolution to 5G will see increase in network complexity Multiple coexisting technologies Network function virtualization Vast differences in terminal capability Significant variation in traffic demand Completely new & varied use cases 2G 5G 4G 3G Complexity 1990s 2000s 2010s 2020s Dealing with opex and network performance in this environment will go beyond the reach of humans

20 Complexity in the connected world e.g. transportation systems are increasingly complex and full of sensors, smart devices and steered by multiple services running on both devices and cloud the integration of components like trucks, trailers and containers shall be simple

21

22 Management and Operations of a Cyber-Physical System Technologies and Components Examples white-label Enterprise Telco presentation application Applied Context Business Business objectives objectives Maximize Optimize User grid utilization Experience Decision Support and Knowledge Management Declarative and Procedural Knowledge Reasoning and Formal Methods Reasoning and Formal Methods phase network drift capacity is X - increasing is appropriate price by for Y youtube for area Z consumption would reduce Knowledge consumption increasing management, average 5%, stabilizing bandwidth verification grid with alternatively and 10% synthesis, would increase reduce reasoning, hydro churn workflow power by optimization 50% by 2% IoT Resource Management Data Analytics Real World Model Analytics Abstraction and Semantics Semantics Semantics and and context context Smart meter for user Cell XXX high overload reactive power outtake Semantic Device annotation, Substation YYY in Norra Entity CELL 2 XXX of IED 0.1 outage Mbit/s modeling, Processed data feeds Average D/M/Y throughput pattern and of average all devices of Sensor in the 5 City: is: Y 1.5 (curve) Mbit/s Stream Real-time analytics, Anomaly Pattern monitoring detected recognition, of cell in Sensor XXX Anomaly shows 8 YYY detection Networks & Devices Sensors & Actuators Observation and Control Raw data feeds Raw data feeds Sensor Cell XXX 5 value = 3.4 = kw22 Data Device2 capture, throughput Sensor Actuator 8 = level 0 commands, kw= 1.1 Mbit/s Cloud1 = Valve 60% 2 processor = 12 load Things and Places Available Available assets assets Cell Smart XXX, meter, Site 4, 10kV Pico 15, Distribution Site Controller Substation, 7, Antenna Energy 3, Service RBS6000 Interface (ESI), Photovoltaic/Hydro, #2, VM1 (cloud1) Electrical vehicle

23 Technologies behind the scenes Expert User Actions Approver Knowledge Data Processing Mission connected things Knowledge Base Generated Knowledge Planning & Reasoning Strategy Online Learning Explanation networks documents Model Repository Data Lake & Streams Predictions &Estimation Deep & Shallow Learning Analysis Property check What-if analysis Ac<ons Raw Data

24 Intelligent Site Power failures detec0on Sleeping cells prediction* Field dispatch prevention Digital Twin - site profiling SLA/ KPI degradation* Prediction accuracy 85% in combined prediction of site down 85% in battery degradation 90% in grid outage 7 out of 10 Sleeping cells correctly predicted up to 24h in advance 9 out of 10 unnecessary site visits predicted correctly At least 1 anomaly detected in 44% of sites over a month KPI accuracy Throughput : 80% Latency : 85%

25 Use Case Sleeping Cell Prediction Description Identify the patterns or triggers that predict the likelihood of a site going to sleeping/silent mode, and automate actions to resolve the issue. Data Requirements PM Counters, Site Parameters (incl. physical locations), Configuration Management, CTUM, Cell Trace and Automation logs CondiGonal Inference Tree Data Models & Algorithms Automations Trigger pre-defined runbooks in an Automation tool to perform remedial tasks (e.g. unlock)

26 Telco Knowledge Graph Telco Knowledge Graph Intelligent Applications Product Documentation Digitized Expert Knowledge Insights Machine Learning Data Mobile Digital Assistant Network Information Contextual Knowledge Structured, Linked Data Learning Models, Rules, Procedures Informed Decision-Making Intelligent Network OperaHons Centre

27 Key insights Data-driven and datacentric research Dealing with heterogeneity though semantics Right data at right time and place Keeping the global state together Mix of MI approaches and techniques ML meets Reasoning Declarative meets Procedural Collaborative Intelligence Frameworks will be needed for success of MI applications Safety Trust Transparency Explainability Privacy

28 Ques%ons & Answers

29