Process Optimization in Manufacturing Plants, by Keyanoush Razavidinani, Digital Business Consultant at A1 Digital.
*Machine Learning School in The Netherlands 2022.
DutchMLSchool 2022 - Process Optimization in Manufacturing Plants
1. A1 Digital ML powered by bigML
Process
Optimization
In Manufacturing
2. Agenda
• Short introduction A1 Digital (and the speaker)
• Digital Transformation
• Necessary to scale beyond traditional
boundaries
• The Pilot Stage - Expectations and timeframe
• Setting the Stage – Asking the right Questions
• Process Optimization in Manufacturing
• The Perfect Granule
• Data Availability
• Features and Labeling
• Learnings & Challenges
A1 Digital powered by bigML
Process
Optimization
In Manufacturing
>3 million M2M Sim Cards
4. Agenda
• Short introduction A1 Digital (and the speaker)
• Digital Transformation
• Necessary to scale beyond traditional
boundaries
• The Pilot Stage - Expectations and timeframe
• Setting the Stage – Asking the right Questions
• Process Optimization in Manufacturing
• The Perfect Granule
• Data Availability
• Features and Labeling
• Learnings & Challenges
A1 Digital powered by bigML
Process
Optimization
In Manufacturing
5. 6
Digital Transformation – A Necessity to Scale the Business
Transformation
The Traditional
Siloed Company
Silo 2 Silo …
Silo 1
Department 1 Department 2 Department …
Models Models
Models
Data Data
Data
Silo
…
Running Pilots in
Departments
Silo
2
Silo
3
Silo
1
Data
Data
Models Models
Data
Data
Common Data Platform
Use
Case 2
Use
Case …
Use
Case 1
Common Libraries
Data
Performance
PoC Stage
Development Stage
Common AI Platform
Common Libraries / ML Models
Data
Agile Teams
APIs
AI Factory
+
6. 7
Digital Transformation – Expectations and Timeframe
A PoC is just the beginning – the least painful and
cheapest part of the transformation
Creating a machine learning model and analyzing
data is just the first step.
Digital operating models require a service-oriented
architecture and new ways of managing and
leadership.
7. 8
Digital Transformation – Expectations and Timeframe
Think beyond the PoC Stage A PoC is just the beginning – the least painful and
cheapest part of the transformation
Creating a machine learning model and analyzing
data is just the first step.
Digital operating models require a service-oriented
architecture and new ways of managing and
leadership.
Value
Scale
Traditional Operating
Model
Digital Operating
Model
Give up
here
8. 9
Ocado – Grocery Delivery McKinsey – Energy Plant Amazon Warehouse
AI Factory - Examples
Ocado built a phenomenal foundation of
data, AI, and robotics
Data, libraries, ML models, and APIs
are shared across the company.
Unrivaled centralized data platform with
details on its products, customers, partners,
supply chain, and delivery environment
AI algorithms in the driver’s seat for
decision-making.
Various models are used for
routing calculations, customer demand
prediction, delivery priorization models,
and much more
Increase efficiencies and reduce
emissions of plant with AI
Multiple years of data
available (temperature, humidity,
internal decisions, outcome)
Plant engineers worked closely
with ML team for guidance
and critical data sources
Deploy >400 ML Models
helping operators make
better decisions and capture
>$60mn value/year
Amazon uses 100s of ML
models in its warehouses.
Visual quality control
Of packages and labels
Prediction analysis of demand
Patterns and item positioning
Packaging size and tape length
Predictive maintenance models
For conveyor belt motors
9. Agenda
• Short introduction A1 Digital (and the speaker)
• Digital Transformation
• Necessary to scale beyond traditional
boundaries
• The Pilot Stage - Expectations and timeframe
• Setting the Stage – Asking the right Questions
• Process Optimization in Manufacturing
• The Perfect Granule
• Data Availability
• Features and Labeling
• Learnings & Challenges
A1 Digital powered by bigML
Process
Optimization
In Manufacturing
10. 11
Decision Framework for Machine Learning – The Checklist
Strategic Framework
Top three
strategic priorities Type of Data
Data Framework
Repitive Data
Storage System
Data Correlation to
Strategic Goals
What is the core service
delivered?
Do network or learning
effects exist?
11. 12
Decision Framework for Machine Learning – The Checklist
Decision Framework
Strategic Framework
Top three
strategic priorities How are decisions made
Existing Software
To make decisions
Human Interpreter
Predictions per Year
Type of Data
Data Framework
Repitive Data
Storage System
Data Correlation to
Strategic Goals
What is the core service
delivered?
Do network or learning
effects exist?
12. 13
Digital Transformation – The Pilot Stage
Silo …
Running Pilots in Departments
Process
Engineer
Silo 3
Manufact
uring
Data - Exoscale Cloud
Models
BigML – Models & API Models
Data
Data
PoC Stage
Foundation and Work Expectations
• Process Chart – Breakdown into parts
• Data Availability
• Existing data in legacy systems
• Value of the features extractable from
existing data
• Existing Work
• Existing models and automations built by
teams
• Value of Automation
• Timeframe
• Which stage is the company in?
• Based on stage, use case, and data
availability a PoC could take form 6 weeks to
6 months
Process engineers, data scientists, and machine
learning experts work together to build the
foundation for the processes, machines, features,
and the data.
13. Agenda
• Short introduction A1 Digital (and the speaker)
• Digital Transformation
• Necessary to scale beyond traditional
boundaries
• The Pilot Stage - Expectations and timeframe
• Setting the Stage – Asking the right Questions
• Process Optimization in Manufacturing
• The Perfect Granule
• Data Availability
• Features and Labeling
• Learnings & Challenges
A1 Digital powered by bigML
Process
Optimization
In Manufacturing
15. 16
The Perfect Granule
• Fluid Bed Dryer Granulator makes
granules from powder
• Powder into chamber
• Constant air stream + fluid spray
• Dry fluid and powder “clumps”
together into granule
• Goal – Optimize process parameters
for granulation process
• Before – Human “feel” vibration to
identify granule progress
Process engineers and machine
learning experts are required to
understand the processes, machines,
features, and the data.
Process Optimization – The Perfect Granule
16. 17
Process Optimization – Data Availability Challenges
16 Process
Parameters are set
and measured
Add sensors to
measure vibration
Granule Quality
data available
Data within legacy systems (data silos) –
need to scrape data
Data is unformatted
Pre-processing is required to extract features
Reduce dimensionality & calculate features
Process engineer expertise required
No vibration measurement available
Check feasibility of vibration analyzis
during process
Accumulate data
Analyze data and discuss with process
engineer
Calculate features
17. 18
Process Optimization – Features and Labeling
Einspritzen?
1805
0 Hz
900
450
1350
04.08. 05.08. 06.08.
Vibration in z-
axis
The combination of process
parameter measurements and
vibration allows to identify the
progress of the granule
Calculate features and label data
based on human experience and
sensor data
18. 19
Process Optimization – Centrifuge Separator
Data
Scraping
Calculate
Features
16 Process
Parameters are set
and measured
Add sensors to
measure vibration
Granule Quality
data available
Cloud
Process Parameters
+ Alarm when
product finished
Data
Processing
Feed
Model
Decision
Output
20. 21
Process Optimization – Lessons Learned
The biggest challenge → managing expectations
Other challenges:
Time management of industry experts
Different data formats used throughout the company
Accumulating the right data from the right places – legacy systems, finding the right
experts within the company
How to position our solution
Digital operating models require new forms of management.
21. AI, cloud, and a common data framework
allow to scale beyond traditional operating
models
Process Optimization
In Manufacturing
Conclusion
Value
Scale
Traditional Operating
Model
Digital Operating
Model
Give up
here
Scalable tools like BigML and the Exoscale
cloud are at the core of the transformation.
22. Contact
A1 Digital (Germany & International)
Keyanoush Razavidinani
Digital Business Consultant
+49 172 1442655
keyanoush.razavidinani@a1.digital