Predictive maintenance - Architecting a Solution with Devices, Services, Big Data and Predictive Analytics

Uli HommanChief ArchitectWW Services

Predictive MaintenanceArchitecting a Solution with Devices, Services, Big Data and Predictive Analytics

ARC301

Marc MercuriArchitect MgrApplied Incubation

Mark KottkeArchitectApplied Incubation

Michael EpprechtArchitectModern Apps CoE

Presented in 2014

Session Objective(s): Describe the predictive maintenance scenario and identify relevant technologies in the MS stack.Define architecture patterns core to the end to end scenario

Session Objectives And Takeaways

Vs. Preventative Maintenance

Connectivity benefits the customer AND the OEM

• Remote Monitoring• Power Grid

• Tolling• Traffic• Navigation

• Safety

3rd Party Services

Road

Vehicles• Social Networking

• Connected Devices• Mobile Network

Operator

• Retail• Insurance• Infotainment

Long term opportunity

Note: Illustrative based on potential one percent savings applied across specific global industry sectors.Source: GE estimates

What if… Potential Performance Gains in Key SectorsIndustry

Aviation

Oil & Gas

Rail

Healthcare

Power

Estimated Value Over 15 Years (Billion nominal US dollars)

Segment Type of Savings

Commercial

Gas-fired Generation

System-wide

Freight

Exploration & Development

1% Fuel Savings

1% Fuel Savings

1% Reduction in System Inefficiency

1% Reduction in System Inefficiency

1% Reduction in Capital Inefficiency

$30B

$66B

$63B

$27B

$90B

Alerts, Analytics, Events and Access

Workflow and Business

Process AutomationThroughput and OEEVisibility, Role-based

and Mobile

Innovative approaches for transformation

People Assets InformationProcess

Productive environments

Sustainable Performance

Reliable process

capabilities

Informed decision making

• Familiar interaction with natural user interfaces

• Simple Role-oriented workspace

• Collaborative communication

• Connected • Remote

monitoring• Business

continuity

• Control• Quality• Standardization• Flexibility

• Secure and timely visibility

• Complete, contextual, accurate

• Predictive and actionable

Monitor, mine, manage pattern1. Monitor and collect events2. Mine system events to produce active model (e.g.

fraud detection, preventative maintenance)3. Manage active event stream via event engine

Event Engine

ModelGeneration

Digital Shoebox

3

21

Data is acquired from devices, sensors, applications and people.

Evaluation, storage, and processing of data is done locally

as appropriate.

If the local implementation is a hub and spoke design, data is communicated to a locally connected hub.

Appropriate data is transmitted to a public or private cloud.

The specific data transmitted, how it is transmitted, and timing of transmission is determined by policy. Policy includes considerations of “three Cs” – context, connectivity, and cost.

Services are utilized to deliver the data, store it, and if appropriate, initiate one or more associated data pipelines.

Compute Storage

Analytic pipelines perform analysis and generate insight. The resulting data delivers one of four things –

• Information for Subscribers• Enhancements of Existing Data• Recommend Action(s)• Initiation of Action(s)

Compute Storage Analytics

Pipeline(s) Insight

Insight is delivered to appropriate human, device, and application audiences in the forms of -

• Alerts/Notifications• Reporting• Command + Control• Data Services• Personalized User Experiences

Compute Storage Analytics

Pipeline(s) Insight

A diversity of

Peer-to-Peer

Device-to-Service Service-to-DeviceMachine-to-Machine communication is non-interactive, automated, and bi-directional information exchange in

operational systems, performed between peers or between satellite systems and their supporting backend services.

Connectivity Patterns

Connectivity considerations•

•

•

•••

•

•

•

••••••

••••

••••

•

•••

•

•

•

•

••

•

Common Activities and Composability

Peer-to-Peer

Device-to-Service Service-to-Device

Service-to-Service

Information Exchange Patterns

Telemetry

Information flowing from a device to other systems for conveying status of device and environment

Inquiries

Requests from devices looking to gather required information or asking to initiate activities

Commands

Commands from other systems to a device or a group of devices to perform specific activities

Notifications

Information flowing from other systems to a device (-group) for conveying status changes in the rest of the world

Telemetry Types•

•

••

••

Telemetry Considerations•••

•••

•••

•••

Signal Characterization••

••

••

••

Policy Considerations••

••

••

The “three Cs” will help determine the appropriate telemetry to deliver at any given time.

Architectural BaselineScale Unit

x10,000 devices

Data Analysis Pipeline(s)

Gateway

Filtering and Aggregation

Routing

Control System

ScaleUnit

ScaleUnit

ScaleUnit

x1,000,000 devices

ScaleUnitDC Boundary

Device Identity

and Metadata

Store

Provisioning System

Data

Ser

vice

s

Gateway Core Architectural Components1. Custom Protocol

Gateway2. Telemetry Pump

and Adapters3. Command

Gateway4. Provisioning

Service and Metadata Store

Windows Azure Service Bus Messaging

Custom Protocol Gateway Host

MQTT CoAP … …

Telemetry/Request Router

Notification/Command Router

Adapters Command API Host

Provisioning Service

Device Metadata and Key

Store

HDIn

sight

BizT

alk

Sv/S

rOr

lean

sAz

ure

Stor

age

SQL

SB

HTTP

HTTP

Devices

AMQP

1

2 3

4

Configs

• Telemetry Adapters take ingress data then adapt and deliver it to raw storage, data stores and/or associated technologies.

• Data Pipelines are workflows that process data with the intent of transformation and/or generation of insight.

• Adapters and Pipelines can also be re-entrant, transforming data and publishing back into it into telemetry processor.

Telemetry Adapters and Data Pipelines

Real-time* Analysis• Observe Telemetry “as it

happens”• React to state changes or trends• React to aggregate observations

• Examples• “device input voltage drops below 11V for more than

3 minutes”• “temperature readings from sensors on this floor

average above 23°C for last 10 minutes”• “sensor failed reporting data for 5 minutes”

• Very short reaction time required

fn

Data-At-Rest Analysis• Mine Telemetry through DB

Queries• Find and track trends or maxima• Analyze expected vs. actual behaviors• React to longer term observations• Hoard for future use

• Variety of Data Store Options• SQL/OLAP• Cassandra, Riak• Hadoop/HDInsight

• Store choice depends on what questions you’d like to ask

flt

Data Lakes

Data Lake

• Data volume and velocity growing• Storage is cheap• With data warehouses… • Designed to answer questions you have today• Elements and attributes not needed often dropped / lost

• Data lakes…• Keep all data for future needs – known/unknown• Includes meta-data tags to help find the data you need later• Feeds data pipelines for downstream needs.. including data

warehouses

Supervised LearningA supervised learning algorithm analyzes labeled training data and produces an inferred Function which can be used for mapping new examples

Training Data• Training data consist of a set of training examples• Each example is a pair consisting of an input object and a desired output

level• As real data evolves/changes, the algorithm can be run against new

training data

Sensor X

Age

What’s the likelihood that a machine will fail soon given device age and data received from sensor X?

MaintInterval

Daily Usage

Supervised Learning Examples What is the ideal maintenance intervalbased on an machine’s daily usage?

Inputobject

Desired

Output

“Not Failing Soon”

Classification

“Failing Soon”Classification

Unsupervised LearningFinding hidden structure when you don’t know the answers.Often finding clusters within the data

Examples• Market segmentation analysis• Organizing computing clusters• Grouping web content, e.g. news stories• Social network analysis X1

X2

Data Pipeline ComponentsData Analysis Pipeline(s)

Data

Ser

vice

s

1 2

3

4

5

6

1. Hadoop2. R3. RDBMS/SQL4. NoSQL5. Storage6. Codename

“Passau”7. ASP.NET Web API

7

Project Passau

• Easy data exploration through web-based interface

• No programming required

• Flexible and extensible

Hadoop Options

•

•

•

Resulting Insight Types••

•

•

••

Command/Control• Tell a device, remotely, to execute a

logical or physical activity• “Give me the status of X” • “Roll 2 feet forward”• “Track this object with the camera”• “Fetch firmware update”

• Remote: Control service, handheld device, etc.

• Latency requirements vary, but often “perceptibly imminent”

Data Services

Deliver Using Open Standards••

•

••••

•

•

•

Consuming Insight•

•

•

•

••

••

Case Study - Altran

Altran Overview• Hi Tech Engineering &

Consulting • $2B 2012 Revenue, 20K

Employees• Aerospace, Auto, Energy,

Life Sciences, Media, Rail, Government

Daily Data Analysis for 50K vehiclesSimulate 1 Month of data Ingest Data @50K msgs/secShow results in Windows 8 dashboard

Data Harvester

Architectural BaselineScale Unit

x10,000 devices

Data Analysis Pipeline(s)

Gateway

Filtering and Aggregation

Routing

Control System

ScaleUnit

ScaleUnit

ScaleUnit

x1,000,000 devices

ScaleUnitDC Boundary

Device Identity

and Metadata

Store

Provisioning System

Data

Ser

vice

s

Device Gateway – Partition Topology

• “Master” manages device provisioning and partition deployment/configuration for all or a well-defined subset of partitions (e.g. one continent)

• “Partition” is a set of resources focused on handling data from a well-defined and known defined device population that has been assigned to and configured into the partition through provisioning. Cross-partition distribution of devices is based on solution-specific logic, allocation within the partition is handled by provisioning.

PartitionMaster

Provisioning API

Provisioning Runtime

Partition Repo

Ingestion Topics Egress

Service Bus Standard Protocol Custom Protocol

Device Repo

Access Control

in0000 inFFFF…in0001 in0002

AMQPS HTTPS MQTT Custom Protocol HostProtocol Adapters

diagall dia

gall diagall dia

gall

Telemetry PumpN Instances

Telemetry Adapter

Telemetry Adapter

Telemetry Adapter

Deployment Runtime out000

0 outFFFF…out0001

out0002

s000 1

s000 2

s03E 7 s000 1

s000 2

s03E 7

s000 1

s000 2

s03E 7

s000 1

s000 2

s03E 7

g0000/rte0000

g0001/rte0000

g0000/rte0001

g0001/rte0001

out0

out1

out2

out0

out1

out2

out0

out1

out2

out0

out1

out2

n Groups of m Routers

SB AMQPS

Altran ArchitectureVehicle Data Simulated Vehicle

Data

HDInsightAzure Blob Storage (ASV) Azure SQL DB

Client Data Services

(Worker Role)

Demo

In Review: Session Objectives And TakeawaysDescribe the predictive maintenance scenario and identify relevant technologies in the MS stack.Define architecture patterns core to the end to end scenario

© 2014 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries.The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.