2. • Intro
• IoT evolution, Big Data in IoT
• Cisco Cloud Platform
• Case Study
• Preventive maintenance of Robots in Automotive industry
• Adaptive, self-learning next-gen Predictive maintenance platform
Agenda
6. Amit: my other passion
Big Data Architect at Cisco Cloud Platform & Services Group
• Big Data Solutions for clients and infra needs using Hadoop, Cassandra
• Analytics platform design
• Data Center infra software abstraction : Firewall as a Service, Networking as a
Service.
Previously Symantec/Verisign, HCL-US, BoA
• Distributed Systems design and implementation
• Hadoop based solutions for large data sets
7. Cisco Cloud Platform
Global platform
deployed across
Cisco and SP
Partners
API-driven, elastic
experience for
developers,
based on open
standards
Cisco-architected
and operated for
rapid application
development and
deployment
10. The Four Eras of Compute
1960 1980 2005 2015
Mainframe
x86 Linux Web VMs iPhone
PC + Web Cloud + Mobile
IoT + Analytics +
Automation (ML)
Cloud Containers
Enterprise Consumer IoT ( Machines )
16. Zero Downtime
• Cisco and Fanuc have created a Zero Downtime Solution (ZDT)
that analyzes data from robots to detect potential problems that
could lead to a failure.
• ZDT is currently used in production with over 6,000 robots at
automotive plants globally. GM alone has deployed ZDT in 27
factories in 5 countries analyzing over 5,000 of robots
• ZDT has successfully detected over 45 cases of potential failure
across 26 production plants over the past year and saved
already customers $40 million
18. Overview
$2 million outage avoided !
Telemetry
collected
Notify robot
manufacturer
and plant
Plant
Data
Collector
Cisco
Cloud
Parts
warehouse
Car plant
Scheduled
maintenance
19. Cisco Cloud
Automotive manufacturer A
Plant 2
Plant 1
Plant Data Collector
Case study: Data Flow
Cisco IoT Platform
Plant 3
Cisco IoT Platform
Cisco BDaaS
ZDT application
Reporting
Analytics
Car manufacturers
Robot manufacturer
Automotive manufacturer B
Plant 2
Plant 1
ZDT Data Collector
Cisco IoT Platform
Plant 3
Notifications
20. Cisco Cloud
Car PlantCar Plant
Batch Layer
Cisco Cloud: High Level Arch Framework
Speed Layer
Serving Layer
Master
dataset
Batch
view Batch
view
Real-time
view
App
Car Plants
Batch
processing
Real-time
view
Real-time
processing
Data
Ingest
Layer
Data
stream
21. Cisco Cloud
Batch
Case study: ZDT Cisco Cloud Pilot details
Real-time
Serving
Master
data
Computed
data
HBase
Ingest
Cisco IoT
Kafka
Flume
Spark Streaming
Batch processing:
Pig, Hive
Impala
ImpalaSQL
schema
Data
API
24. Why Predictive?
Car Production facilities operate at high volume
Unexpected downtime creates considerable losses
There is a need to be informed of a potential robot, controller or process problem
before unexpected downtime occurs
Early detection is key in the following scenarios
Mechanical failures
Process control failures
System issues: Controller
Maintenance reminders
Not-too-early and not-too-late detection is “key”
Too early is expensive in the long run
Too late is detrimental as well
Finding the sweet spot is key to the most “optimal solution”
27. Predictive analytics: Increased ROI
Report
• What happened
Analyze
• Why did it
happen
Monitor
• What is
happening now
Predict
• What might
happen
Increasing ROI and Complexity
28. Data Modeling details
Initial Dataset
Run/Evaluate
Models
Gather Data
Define
Problem
Validation Dataset
Test
Model
Select
Model
Test Dataset
Apply Model
Run Prediction
29. Stream Processing Layer
HDFS
Data Ingest Layer
Predictive Analytics: High level architecture
Learning Layer
Action Layer
Raw
dataset
Processed
dataset
Kafka
Cisco
IoT
Platform
Near “real-time” (micro-batch)
processing
( Spark )
Machine Learning
( Spark ML )
HDFS
Knowledge
Base
Operational
Dashboard
platform
( custom built /
Sensu
customized )
30. Stream Processing Layer
HDFS
Data Ingest Layer
Predictive Analytics: High level architecture
Learning Layer
Action Layer
Raw
dataset
Processed
dataset
Near “real-time” (micro-batch)
processing
( Spark )
Machine Learning
( Spark ML )
HDFS
Knowledge Base
31. File Formats: Avro vs Parquet vs ORC
Avro is row-based
storage format,
optimized for scans
of all fields in a row
for each query
Parquet is column-
based, best used
when dataset has
many columns and
only a few columns
are worked on
ORC is column-
based as well
32. Spark based Predictive platform on Hadoop
Data Integration ( Kafka, Sqoop, Flume )
Storage for any type of data
Filesystem
(HDFS)
Online NoSQL
(HBase)
Workload Management ( YARN )
Machine Learning
(Spark, Mahout)
Stream Processing
(Spark)
33. Stream Processing LayerData Ingest Layer
Predictive Analytics: High level architecture
Learning Layer
Action Layer
Operational
Dashboard
platform
( custom built /
Sensu
customized )
34. Action Layer
Predictive Analytics: Action layer
Event store
Event consumer
API based event
Topic
Consumer for
email
Dashboard
middle-tier
API for Ad-hoc
queries
Consumer for
PagerDuty
Custom built /
Sensu
customized
FANUC provides robots with high reliability and uptime averaging 9+ years (80k hours) mean time between failures.
But this was not enough, because they know that when a production line goes down unexpectedly it can cost a company as much as $20,000 per minute or and $2 million for a single incident.
In a factory that’s kicking out one vehicle per minute, every minute of stalled production is hemorrhaging profits, labor expenses and more.
Zero Downtime (zdt)
Reacting to unexpected downtime can mean an extended period of time for:
problem solving
ordering and receiving parts that are needed
scheduling a service person
servicing the robot and/or system
Production facilities are operating at high volume
Single robot failure can stop the line
Unexpected downtime creates considerable consequences
Need to be informed of a potential robot, controller or process problem before unexpected downtime occurs
ZDT is FANUC's new diagnostic tool that detects critical information about the robot's mechanical, maintenance and process/system health to alert customers about potential system or product issues.
Data relevant for the maintenance issue is sent to the Cisco Cloud where the cloud analytics engine captures the “out of range” exceptions and predicts the maintenance need.
Then, an alert is sent from the cloud application to FANUC service personnel and to the manufacturing customer about the need for replacement part.
The part is then shipped to arrive at the factory in time for the next scheduled planned maintenance window.
Anything on the robot controller that predicts or prevents robot downtime is part of ZDT
early detection of impending
Mechanical failures
Process control failures
System issues
FANUC Current installation base:
250,000 robots
2.4 million CNCs
12.7 million servo motors
Zero Downtime (zdt)
Reacting to unexpected downtime can mean an extended period of time for:
problem solving
ordering and receiving parts that are needed
scheduling a service person
servicing the robot and/or system
Production facilities are operating at high volume
Single robot failure can stop the line
Unexpected downtime creates considerable consequences
Need to be informed of a potential robot, controller or process problem before unexpected downtime occurs
PaaS: Develop fog applications. Our first IoT PaaS offering, called Cisco DSX, simplifies fog application
development in several ways:
◦ Device abstraction: Fog applications need to communicate with many types of IoT devices. Creating a
separate application for each vendor’s temperature sensor, for example, would be impractical. Cisco
DSX saves application developers this effort by providing an abstracted view of IoT devices.
◦ Support for multiple development environments. IoT applications that deliver machine as a service
(MaaS) are typically developed in various environments and programming languages. With Cisco DSX,
fog nodes can support multiple development environments.
◦ Simplified management of fog applications. Managing a growing number of fog applications would also
be impractical. Cisco DSX simplifies management and automates policy enforcement.
SPG Sofwqare Platform Group
Motion/Mechanical Health
Examples: Motor and Reducer deterioration, gear backlash , excessive servo-off events, Dislocation prediction for delta robots, etc.
Process Health
Examples: Air regulator failure, canister high torque, ServoGun motor/drive deterioration, abnormal tip wear, gas flaw, wire spool low, part rate increase beyond, etc.
System Health
Examples: Memory utilization, Ethernet Transmit/receive errors, Vision performance, CPU performance issues, etc.