TITAN Group is an international cement and building materials producer aspiring to serve the needs of society, while contributing to sustainable growth with responsibility and integrity. During the era of Industry 4.0 Titan Group is investing in a Digital Transformation strategy with several Data Science projects being part of it. Alexandros Tsolkas will introduce us to Titan Group and its activities and he will give an overview on how we apply Data Science, from the data management to the way we collaborate with the industry experts. Panagiotis Ypsilantis will summarize the Supply Chain Advanced Analytics use cases in the cement industry (Demand Forecasting, Supply Network Optimization, Inventory Optimization) that currently are developed in Titan and he will present with more details the way that Titan optimizes its Spare Parts Inventory using forecasting and Monte Carlo simulation models. Finally, Ilias Panagoulias is going to describe the challenges and the results of the implementation of a Real Time Optimization platform in the cement process.
2. Titan group
Who we are
TITAN Group is
an international
cement and
building
materials
producer
Founded in 1902
Listed on the ASE since 1912
14 cement plants in 10 countries
5,400+ employees
2 2nd PyData Piraeus – October 22nd ,2019
3. Titan group
What we do
We supply the materials
to build structures and
infrastructures which,
in turn, provide shelter,
enable commerce and
foster connectivity
Cement
Ready-Mix
Concrete
Aggregate
s
Fly
ash
Building
blocks
Waste
management
and
alternative fuels
19.2 m MT
5.6 m m3
16.0 m MT
0.32 m MT
3 2nd PyData Piraeus – October 22nd ,2019
4. Titan group
Where we operate
Our diversified
portfolio of assets:
14 cement plants
in ten countries
across five
continents
14 cement plants: Albania 1 ● Bulgaria 1 ● Egypt 2 ● North Macedonia 1 ● Greece 3 ● Kosovo 1 ● Serbia 1 ● Turkey 1 ● USA 2 ●
Brazil 1
Other assets include grinding plants, distribution terminals, ready mix plants, quarries
Key Terminals
Cement Plants
4 2nd PyData Piraeus – October 22nd ,2019
5. How should a risk-averse, centenarian, heavy-industry company in
a slow-moving sector, think about this new world?
5 2nd PyData Piraeus – October 22nd ,2019
7. What could it mean for cement?
The dawn of Industry 4.0
CUSTOMERS
PLANNING/SALES CEMENT PLANT
SUPPLIERS
LOGISTICS
LOGISTICS
Average
cement plant
generates >1
TB of data p.a
BIM/Smart
Buildings
Fleet
monitoring &
optimization
Data driven
Demand
forecasting
Assets’
predictive
maintenance Inventory
optimization
Supply network
optimization
Example of potential artificial intelligence implementations
Real time
customer
experience
Assets’
optimization
7 2nd PyData Piraeus – October 22nd ,2019
9. Our test & learn approach
“Test and Learn”
“Scale up”
Scale-up in all
BUs
Roll-out
successful
initiatives
Implement across
different areas of
activity
Align with TITAN’s
strategy
Experiment with many
pilots: No regret moves
• Verify impact &
implementation
requirements
A
Build digital capabilities
& infrastructure
• Acquire digital talent
• Monitor market &
potential partnerships
B
“Capture impact
across areas”
9 2nd PyData Piraeus – October 22nd ,2019
10. Upgrading our Infrastructure & Data Management
Process
Instrumentation
Emmissions
Instrumentation
LAB
Instrumentation
QCX
Quality
Data
Plant
sensors
100 signals
1200 signals
150 signals
SCADA
Servers
Gateway
VIRTUAL
Servers
PLC Real Time
reporting
Use of analytic &
machine learning tools
Model
development
Plant’s users HQs
Plant Edge
storage
& ONLINE
computing
ü Data flow to SAP
ü Shift & Executive
reports
ü Mobile App for KPI
Data Cloud
5 years data per 1s
An average
cement plant
creates >1.0 TB
of data annually
10 2nd PyData Piraeus – October 22nd ,2019
11. Upgrading our Infrastructure & Data Management
1. Diagnostic of our current SENSORS
INFRASTRUCTURE
• prerequisite for digital projects
(sensor data for algorithm
development)
• guide for future investments
(“best practice”: number, type &
setup of sensors)
2. Comprehensive CYBERSECURITY PLAN to connect
the Plants’ process control network with the
corporate network and the outside world
11 2nd PyData Piraeus – October 22nd ,2019
14. Network Optimization
Objectives
Identify most profitable product flow
from plants to customers
Improve service level while
reducing costs
Increase asset utilization
14 2nd PyData Piraeus – October 22nd ,2019
15. Network Optimization
Implementation in Titan America – Optimize current network
15
1
2
3
4
5
Plant 1
Terminal 1
1
2
3
4
5
Plant 1
Terminal 1
Optimize network flows
- Identify which customer is
best served by which plant
~ 3%
Reduction in
Logistics Costs
Zoom In
TA Cement Plants
TA RMC Plants
Customers
2nd PyData Piraeus – October 22nd ,2019
17. Demand Forecasting
Objectives
Better Production planning – be ready for peaks in
Demand
On time orders of Raw Materials (especially ones
with high lead time)
Optimal schedule of Maintenance outages (target
low demand seasons)
Benefits of
forecasting
Demand
accurately
17 2nd PyData Piraeus – October 22nd ,2019
18. Demand Forecasting
Triple exponential smoothing – Implementation in R
Forecast Accuracy
Comparison of actual vs. forecasted values
In Industry, Forecast Accuracy >60% is considered as adequate.
Next Step?
Forecasting period
Predictive Intervals:
§ Best Case Scenario
§ Worst Case Scenario
Real values
Fitted values
Forecast values
Train Test Forecast
18 2nd PyData Piraeus – October 22nd ,2019
21. Spare parts Inventory Optimization
What is it about?
Spare Parts are many
>12,000in one plant alone
(too many…)
Spare Parts are NOT
Consumables
It is not straightforward to calculate
their rate of consumption
Typical inventory policy
min – max Order when stock
reaches min level
But…
We must set min carefully so that
• We don’t keep too much stock
• We don’t run out of parts while we wait for
the delivery of our order
21 2nd PyData Piraeus – October 22nd ,2019
22. Inventory Optimization using Advanced Analytics
How is it done?
• Data extraction (SAP)
• Transformation (R)
• Data validation
• Segmentation
Data
Ingestion
22 2nd PyData Piraeus – October 22nd ,2019
23. Spare Parts Segmentation
Different inventory policy per segment
1.Consumption
Frequency
2. Demand Volatility 3. Lead Times
DC E
LOW RLT
(Max RLT < Min IDT)
LOW VARIABILITY
(Consumption Qty)
NO CONSUMPTION
1 CONSUMPTION IN ALL
HISTORY
END OF LIFE?
(used to have
consumption)
≤1 CONSUMPTIONS /
YEAR
>1 CONSUMPTION /
YEAR
HIGH RLT
(Max RLT ≥ Min IDT)
HIGH VARIABILITY
(Consumption Qty)
12K Spare Parts
11 2
BA
2
4. Material Criticality
23 2nd PyData Piraeus – October 22nd ,2019
24. Inventory Optimization using Advanced Analytics
How is it done?
• Data extraction (SAP)
• Transformation (R)
• Data validation
• Segmentation
Data
Ingestion
Consumption
Distribution
Lead Time
Distribution
Consumption
over
Lead Time
Distribution
Distribution
Fitting
24 2nd PyData Piraeus – October 22nd ,2019
25. Addressing min-max using Advanced Analytics
Typical inventory
policy min – max Order when stock
reaches min level
But…
We must set min carefully so that
• We don’t keep too much stock
• We don’t run out of parts while we wait for
the delivery of our order
Consumption
Distribution
Lead Time
Distribution
Consumption
over
Lead Time
Distribution
25 2nd PyData Piraeus – October 22nd ,2019
26. Inventory Optimization using Advanced Analytics
How is it done?
• Data extraction (SAP)
• Transformation (R)
• Data validation
• Segmentation
Consumption
Distribution
Lead Time
Distribution
Consumption
over
Lead Time
Distribution
Data
Ingestion
Distribution
Fitting
• Definition of cost function
• Target service level
• Monte-Carlo Simulation
Inventory
Optimization
26 2nd PyData Piraeus – October 22nd ,2019
27. Fact-based solution
Optimize target function
Algorithm will define the inventory policy that
minimizes the cost function
Inventory Holding Cost
Cost of not having the
part when required
27 2nd PyData Piraeus – October 22nd ,2019
28. 2nd PyData Piraeus – October 22nd ,201928
Inventory Optimization
Want to have a look under the hood?
1. Simulating the “real” process
Event based simulation consisting of
consumptions, order placements, material
receipts events
4. .. in order to optimize the
policies.
Running thousand of instances on hundreds
of scenarios to identify the policy with the
optimal cost that satisfies our service level
constraints
3. .. and purchases
from request time .. to order creation.. to
material delivery events
2. .. by simulating consumptions
with detailed inter-demand times and
consumption quantities
30. The cement production process
Cement Plant simple process diagram
1. Raw Mill is the equipment used
to grind raw materials into
“rawmix" during the
manufacture of cement
2. Rawmix is then fed to a Kiln,
which transforms it into clinker
3. The Cement Mill grinds the
hard, nodular clinker from
the cement kiln into the fine
grey powder that is cement
30 2nd PyData Piraeus – October 22nd ,2019
31. Optimization of Vertical Mill
Objectives
Key Targets:VRM Optimization
1.Maintain Quality: Minimize standard
deviation of quality KPIs
2.Throughput (Feed Rate): Increase mill
productivity in tons/hour
3.Energy: Minimize specific energy cost
for given throughput
Maximize production
Maintain quality
(constraint)
Minimize
specific
energy cost 1.Quality: Maintain material fineness
standard deviation at target levels
2.Throughput: Maintain and ideally
reduce the unscheduled shut downs
due to operational reasons
3.Energy: Maintain or increase mill
operating time during off-peak hours
(with lower energy cost), minimize the
operation of mill during peak hours
Constraints to be considered:
Quality Energy
Throughput
31 2nd PyData Piraeus – October 22nd ,2019
32. Optimization of Vertical Mill
Composite Model design
An RTO should be able to suggest at specified time the
values of the manipulated variables that
maximize/minimize our target function keeping the
operational constraints that the plant has set
In this optimization problem we use machine learning
in order to predict the outcome of key variables
according to given operating conditions
Manipulated
Variables
Informative
Variables
Constraints
Target Function
32 2nd PyData Piraeus – October 22nd ,2019
33. Optimization of Vertical Mill
How does it work: use of AI in a machine learning system
33 2nd PyData Piraeus – October 22nd ,2019
34. Optimization of Vertical Mill
• On-site diagnostics (define problem)
• Data capturing, structuring and cleaning
Preparation &
data validation
• Data analysis, optimizer model design
• Simulation (lab phase) & impact
assessment
Proof of concept
• Test & calibrate systemOpen loop trial
• System operation
Close loop
(commissioning)
Screenshot of a Control Room Operator screen for the Vertical Raw Mill
Project implementation steps / methodology
34 2nd PyData Piraeus – October 22nd ,2019
35. Challenges building an RTO
How a ML algorithm learns from noisy data?
Most sensor data are noisy variables with
high SD even on stable operating
conditions.
Appropriate data preprocessing is needed
in order to smooth the data without
loosing important information.
35 2nd PyData Piraeus – October 22nd ,2019
36. Challenges building an RTO
Do we really need data on such a high granularity?
Vibrations can cause mill stoppages
resulting in high downtimes.
Vibrations can occur in less than a minute,
it is crucial an RTO to be able to predict
and avoid them.
36 2nd PyData Piraeus – October 22nd ,2019
37. Challenges building an RTO
Are your data reliable?
Sensors may malfunction at spontaneous times or need maintenance and recalibration.
Data quality checks should be done not only before model training but also when RTO is in operation.
37 2nd PyData Piraeus – October 22nd ,2019
38. Challenges building an RTO
Remove the outliers! Or not?
Outliers usually can harm you ML
algorithm.
However a plant operates most of the
time in the same conditions generating
data in a specific space.
Can these extreme cases help your
algorithm learn the real relationships
between the variables or they are
abnormal operating conditions that you
cannot model?
38 2nd PyData Piraeus – October 22nd ,2019
39. Challenges building an RTO
How do you handle lab measurements?
Blaine and Fineness are the most
important quality characteristics of the
end product.
Blaine and Fineness are measured in the
lab from samples taken from the mill
usually every 1-2 hours.
39 2nd PyData Piraeus – October 22nd ,2019
40. Challenges building an RTO
Are the correlations you observe correct?
40 2nd PyData Piraeus – October 22nd ,2019
41. Challenges building an RTO
Synchronize your signals!
The material we put on the mill needs
significant time to become end product.
E.g. the blaine measurement of a sample
we collect at time t is a result of the feed
rate at time t-n.
It is important to estimate as accurately as
possible these time delays om the
variables in order to get meaningful
correlation between them.
41 2nd PyData Piraeus – October 22nd ,2019
42. Challenges building an RTO
Optimal vs fast solution
The RTO is designed to provide values for the manipulated variables every 30
seconds.
The choice of the ML and the optimization algorithm is
done taking into account this constrain
An ensemble model may give accurate results but an
MLP can make predictions really fast.
A genetic algorithm can avoid local maxima but brute
force on a constrained search space may be also
sufficient.
42 2nd PyData Piraeus – October 22nd ,2019
43. Challenges building an RTO
Summary
• We do spent 90% of our time
cleaning and preparing our data
• We use and test several approaches
but we select the one that satisfy the
business needs
• We collaborate closely with our
automation engineers and process
experts
We already have installed
RTOs in the plants on USA
& Brazil
43 2nd PyData Piraeus – October 22nd ,2019