iiot : old wine in a new bottle?
TRANSCRIPT
Predictive Analytics : Why (I)IoT is different
!Venu Vasudevan, PhD!
!Next.io (Consultant IoT | Big Data)!Adjunct Professor, ECE, Rice U.!
@venuv62!
Me
Intrapreneur. balanced diet of IoT & predictive analytics
๏ IIoT for asset management. Key contributions to Zigbee!
๏ Shazam for IoT - IoT accessory Home/Auto!
๏ Iridium predictive fault management!
1Mill measurands/sec. then satellite ~ now thermostat!
๏ Predictive video analytics (acquired by WatchWith)!
Agenda: Predictive & IIoT
• Why in the limelight?!
• Now. is it new-and-unique or sum-of-parts!
• Next. will it be new-and-unique or sum-of-parts!
IIoT Market Potential
$150B addressable market!by 2020!
Low(er) business friction!- IIoT Technology creators !
are also customers!
Predictive ability : Mandatory, not optional
over-doing!processes!expensive!
under-doing!processes!
catastrophic!
rightsizing a!dynamic, predictive process!
(time | business context)!
e.g. too much ‘routine’ !maintenance. lightly used !
equipment!
e.g. not enough!maintenance. !
high risk equipment!
Business Focus : from reliability to optimization
Predictive Analytics : IoT Challenge
Sources. ParStream,IBM IoT surveys!
Spot
ty D
ata!
‘Goo
d’ p
redi
ctio
ns!
Predictive Analytics : IoT Challenge + Opportunity
high quality,!high velocity!predictions!
with incomplete, untidy data!
Source. Keystone Strategy!
Spot
ty D
ata!
‘Goo
d’ p
redi
ctio
ns!
long runway for predictive!
Challenge : Data-Insight Gap
• There is no ‘free lunch’ : better predictions need more data!
• Ways to narrow the gap!
• (Volume, Velocity) faster, fatter path from data to decisioning!
• (Variability) clever ways to clean data at scale!
• Match best algorithm for the data at hand!
data maturity!
insi
ght!
insight !aspiration!
data !reality!
variability!volume! velocity!
The ‘gap’ is not unique to IIoT. The reasons for it are ..!
IIoT vs Consumer Web : Same gap, different reasons
Consumer IIoT
Capture Hard!(consumers don’t cooperate)!
Easy!(‘things’ always
cooperate - for a price)!
Sanitization Medium!(simpler data types)!
Hard!(gnarlier data types)!
Modeling & Integration
Easy!(e.g. eyeballs, dwell time)!
Hard!(complex data models)!
IIoT+Predictive:more than sum of parts?
IoT!
Predictive!Analytics!
retrospective! descriptive! prescriptive!predictive!
What’s the current IIoT+Predictive architecture?!Does it address the data-insight gap?!
What architectural changes would close the gap?!
depth of insight!
scale!
Now : Cloud-Centric (I)IoT architecture
collect!
learn!
act!
sense!
store.query.!
analyze.predict!
automated | human!
capture.filter.!
cloudedge
scale
scale
Next : Edge-heavy IIoT architecture
collect!
learn!
act!
sense!
store.query.!
analyze.predict!
automated | human!
capture.filter.!
edge edge
cloud
responsiveness
scale
Sensing Data Challenge
Option1. data goes to decisioning !Fatter, faster pipes!
Continuous flow!
Option 2. decisioning goes to data !Intelligent Edge !Periodic updates!
sense!
getting data and decisioning together!
Edges make IIoT Faster
GE Blog - Edge: A Door to the Data Kingdom!
➡ Edges distribute predictive services (cloud vs edge)!➡ policy vs behavior !➡ long-term vs real-time !
➡ architectures for flexible (re)distribution of predictive decision logic?!
Edges make IIoT Faster and Cheaper
➡ Edges distribute predictive services (cloud vs edge)!➡ policy vs behavior !➡ long-term vs real-time!
➡ how will predictive decision logic move to where the data is?!
Jasper. The hidden costs of delivering IoT!
Slow lakes to fast streams
• Now. Transition from data lakes to data streams!
‣ 30-100x speed up : streams over lakes!
‣ needed to deal with real-time IIoT traffic!
‣ lambda architectures balance prediction speed and accuracy!
• Next ….!
untidydata
firehose
cleananalytics
fast & good
slower & much better
Lambdaarchitecture
collect!
Hadoop!
Spark!
Edge Filtering : Slimming diet for fat streams
fitting predictive decisioning logic fit in super-small footprints!
Opportunity : Machine Learning at unprecedented scale
• Machine-learning-as-a-service - rich set of algorithms, solution templates - immediate impact in: !
• problems with established procedures!
• and clean data!
Source. Cortana Intelligence Gallery
learn!
Challenge : Clean Data
• State-of-the-art ML — promises dramatic improvement. But ‘clean data’ hungry!
• Deep Learning 3x better than Regression for electricity demand forecasting!
• needs 1.5 million data points for training (over 4.5 years)!
• Limiting factor is the data quality !
data maturity!
insi
ght!
insight !aspiration!
data !reality!
variability!volume! veracity!
Stanford study. Electricity demand forecasting. Deep learning 3x better than ‘classic’ m/c learning!
Challenge : Clean Data
• State-of-the-art ML — promises dramatic improvement. But ‘clean data’ hungry!
• Deep Learning 3x better than Regression for electricity demand forecasting!
• needs 1.5 million data points for training!
• Limiting factor is the data quality!
Source. HP Enterprise Labs study!
Training Data Training Time
(IoT) signals
3 million frames! days!
Vision 14 million images!
3 days w/ 16000 cores!
2-Tiered Machine Learning for IIoT
• Intelligent IoT data cleansing layer (e.g. Bitstew) - Machine Learning turns dirty data into clean data!
• low-level data cleaning pushed to the edge!
• semantic integration between data sources in the cloud!
• Predictive Layer - Machine Learning turns clean data into clean insights!
interfaces between cleansing & prediction? !
Conclusion
Present : Cloudy
• embrace. leverage cutting edge cloud and ML services!
• extend. adapt to IIoT business processes!
Future : Edgy
• hyper decentralized intelligence and data!
• systems that understand ‘normal’ and ‘deviation’!
• predictive systems that have both response velocity and depth of insight!
Questions?
[email protected] @venuv62
Predictive Analytics : IoT Challenge
good enough,!high velocity!predictions!
with incomplete, untidy data!(hourglass - with decay
statistic)!
Source. Par stream IoT survey!
Challenge : Clean Data
• State-of-the-art ML — promises dramatic improvement. But ‘clean data’ hungry!
• Deep Learning 3x better than Regression for electricity demand forecasting!
• needs 1.5 million data points for training (over 4.5 years)!
• Limiting factor is the data quality !
Stanford study. Electricity demand forecasting. Deep learning 3x better than ‘classic’ m/c learning!
Fast Accurate Clear
Naive Bayes Yes! Low! Somewhat!
Regression Yes! Medium! Yes!
Decision Trees Yes! Medium! Somewhat!
Deep Learning No! High! Heck no!