This presentation given by Think Big's senior data scientist Eliano Marques at Digital Natives conference in Berlin, Germany (November 2015), details how to go from experimentation to productionization for a predictive maintenance use case.

1. ​From insights to production with Big Data Analytics
​Eliano Marques – Senior Data Scientist
​November 2015

3. Large scale solutions typically are part of a discovery
process and fully integrated with the organization strategy
Big Data Analytics Strategy and Ambition
1
Business analytics roadmap
Capture of analytics use
cases and development of
analytics roadmap(s) with
business areas
Productionisation
Large scale
deployment of
analytics use case
based on agile scrum
principles & methods
Analytics
1
23
4
Experimentation
Agile analytics discovery PoC
on offline/ online data to
prove analytics potential prior
to decision on large scale
productionisation
Validation
Decision on
whether to promote
analytics use case for
productionisation
Shared Big Data Analytics governance

4. Use case – Predictive Maintenance
Business analytics roadmap
CFO & Director of
Assets/Production
• What is the outcome of different capital investment for the next 5
years? How do I measure the impact on maintenance?
• Which assets/parts should be targeted for replacement? How to
prioritise them over time?
• How to plan ahead overall costs? What options are available?Director of
Operations
• How to predict demand for reactive maintenance? Can it be
reduced? What is the optimal mix between pro-­active vs. reactive
maintenance?
• How to predict stock levels for assets/parts? Can it be minimise?
• What capacity is needed? Do we need to sub-­contract?
Field Teams
Lead
• How to increase field force efficiency? How can we reduce
engineering visits?
• How to prioritise faults?
• How to predict false alerts?
Strategy
Tactical
Operational
1

5. Use case – Predictive Maintenance
Experimentation
Production Team
Experiment Owner
Business
and
data
Workshops
Experiment
Development
Experiment
Testing
Experiment
Results
Key
activities:
Key
iterations:
Who’s
involved:
Weekly sessions to check
experiment progress and
validate initial results
Delivery workshop with
program management to
share experiment results
Initial workshops between
experiment owners, data
owners, data engineers and
data scientists
Data engineers
Data Scientists
Key
Outputs:
H1: What's the impact of different
capital investment strategies?
H2: Can sensor data be use to predict
time-­to-­fail or risk-­to-­fail of asset parts?
H3: How to minimise faults detection
root-­cause and uplift efficiency?
• Segment field force by
time to detect root
cause patterns
• Predict root-­cause of
failure by type of
asset/part
• Validate/test models with
key stakeholders
• Link sensors with faults
• Prioritise sensors by
criticality of failure
• Develop models and
Predict time/risk to fail by
asset/part
• Validate/test models
with key stakeholders
• Build target investment
models linked with
maintenance, volumes
and workforce
• Develop simulation
tool and run scenarios
on demand
• Validate/test solution
with key stakeholders
2

6. Use case – Predictive Maintenance
Validation
Business
case
assumptions
Business
case
development
Workshop
preparation
Validation
workshop
Key
activities:
Key
iterations:
Who’s
involved:
Meeting with business area
lead to validate business
case
Validation workshop with
steering committee to obtain
approval for moving solution
to production
Meetings with production
team and business area
leads to get business case
inputs
Key
Outputs:
H2: Can sensor data be use to predict
time-­to-­fail or risk-­to-­fail of asset parts?
Pos-­experimentation question:
Is it worth moving to production?
Experiment team
Experiment Owner
Steering Comm.
Production team
Analytics
Technology
costs
and
changes
assumptions
Business
value
assumptions
Business
case
Downstream
ApplicationsInformation Sources
Evaluate
Source
Data
Prepare Source
Metadata
Prepare Datafor
Ingest
Enterprise Data Lake
Sequence Automate
Apply Structure
Compress Protect
DashboardEngine
Collect & Manage
Metadata
Perimeter-Authentication-Authorisation
Ingest
3
• New ingestions? How
many models? Prediction
frequency? Rules
engine?
• How users will access
and make decisions on
demand?
• What’s the size of
benefit? Is it tangible?
• Is the use case viable
financially? What’s the
ROI? What’s is the Pay-­
back period?

7. Use case – Predictive Maintenance
Productionisation
Release
Planning
Create
Project
Backlog
Production
Deployment
Key
activities:
Key
iterations:
Who’s
involved:
Bi-­weekly sign-­off of development
progress by program management
and business area lead
Regular meetings in an agile
scrum format including sprint
planning, daily scrums, and
sprint review
Key
Outputs:
Experiment team
Experiment Owner
Production Team
Scrum Master
Gov.,
Maint &
Training
H2: Can sensor data be use to predict
time-­to-­fail or risk-­to-­fail of asset parts?
Pos-­experimentation question:
Is it worth moving to production?
YES
Sprint
Cycles
Model
3
Model
2
Model
1
• Business and field
engineers can now act on
real time signals based on
predictions of time/risk to
fail for assets and parts
• Rules can be automated
to act on high-­risk threads
• Pro-­active maintenance
decisions can now be
made to optimise costs
and maintenance
efficiency
Downstream
ApplicationsInformation Sources
Evaluate
Source
Data
Prepare Source
Metadata
Prepare Datafor
Ingest
Enterprise Data Lake
Sequence Automate
Apply Structure
Compress Protect
DashboardEngine
Collect & Manage
Metadata
Perimeter-Authentication-Authorisation
Ingest
Solution running
4
✔

8. Think
you
Thank Big