3. Personal Background
Former Deputy CEO for Finance and Economy at
Magnitogorsk Iron & Steel works (MMK)
Former non-executive director, member of the Committee for
Strategic Planning in the Board of Directors of MMK
Adviser to the Chairman of the State Corporation "Bank for
Development and Foreign Economic Affairs”.
Technological outlook from financial standpoint
4. Manufacturing IT spending is often underestimated
4 Source: IDC #257386, Worldwide Vertical Markets IT Spending 2014-2019 Forecast.
5. But it doesn’t always bring added value
MES / ERP
Business process automation (e.g. accounting)
Infrastructure and support
5
6. Though everyone knows what is a priority
6 Source: KPMG CIO Survey 2016 https://assets.kpmg.com/content/dam/kpmg/xx/pdf/2016/10/cio-survey-2016-manufacturing.pdf/
Compared to the all-industries average,
manufacturing companies place a higher
priority on:
〉improving business processes (72% vs.
56%)
〉increasing operational efficiencies
(65% vs. 57%)
〉saving costs (60% vs. 50%)
What are the key business issues that your management are looking for IT to address?
7. Where to focus?
Core manufacturing processes
Less or no capital spending
Clearly measurable value
7
And that is exactly where AI plays its role well!
8. “
” | Jeff Bezos, Amazon
The outside world can push you into Day 2 if you won’t or can’t embrace powerful
trends quickly. If you fight them, you’re probably fighting the future. Embrace
them and you have a tailwind.
These big trends are not that hard to spot (they get talked and written about a lot),
but they can be strangely hard for large organizations to embrace. We’re in the
middle of an obvious one right now: machine learning and artificial intelligence.
Over the past decades computers have broadly automated tasks that
programmers could describe with clear rules and algorithms. Modern machine
learning techniques now allow us to do the same for tasks where describing the
precise rules is much harder.
8
9. “
” | Jeff Bezos, Amazon
But much of what we do with machine learning happens beneath the
surface. Machine learning drives our algorithms for demand forecasting,
product search ranking, product and deals recommendations,
merchandising placements, fraud detection, translations, and much
more. Though less visible, much of the impact of machine learning will be
of this type – quietly but meaningfully improving core operations.
9
17. What is AI for manufacturing?
Complex algorithms that accomplish tasks by
themselves instead of being explicitly
programmed
Learning from past historical data on specific
plant and equipment
Predictive and prescriptive capabilities
Models that deliver best
operational decisions for
important processes in real-time
18. In an ideal world
18
Output parameters may be precisely defined by a described dependency
L(z)
0
z
Technological
process
Input
parameters
Output
parameters
Process parameters
19. In reality
19
Output parameters have permitted ranges, since they cannot be precisely predicted due
to fluctuations in the inputs and process itself
Technological
process
Input
parameters
Output
parameters
Process parameters
L(z)
0
z
21. Optimization of ferroalloy use
What: recommender service to
minimize use of ferroalloys
When: July 2015 – June 2016
Where: Moscow – Magnitogorsk
Who: Magnitogorsk Iron & Steel
Works, Yandex Data Factory
Results: 5% average decrease, >$4m
yearly expected savings
22. Before we start
Steel
An iron-based alloy; chemical requirements strictly specified
The final product
Pig iron
A high-carbon iron-based alloy; its chemistry is known
Contributes 3 /4 of steel raw materials
Scrap
Recycled material; its chemistry is poorly known
Contributes 1 /4 of steel raw materials
Ferroalloys
Iron-based alloys; their chemistry is known
The additives providing the required chemistry of the steel22
23. How the process looks?
23
Oxygen
furnace
Blast
furnace
Ladle
treatment
Slabs
Scrap
Ferroalloys
Steel
Scrap
24. How the process looks?
24
Oxygen
furnace
Blast
furnace
Ladle
treatment
Slabs
Scrap
Ferroalloys
Steel
Scrap
25. Problem statement
Keep the use of ferroalloys to a minimum
during each smelting
Make sure that the resulting chemical
composition of steel complies with the
specification
Currently the decision is made by an
operator based on a set of rules and
individual expertise
33. Optimization potential
Often we get somewhere in the
middle of the required range
But we want to consistently get
closer to the lower boundary
34. Why we chose this project?
34
Significant cost cutting
expected
Yandex Data Factory as a
trusted partner, spin-off of
leading IT company
Team on the shop floor
willing to cooperate and
experiment
Enough data of seemingly
high quality
35. Data used in the project
Historical data on more than 200,000 smeltings, including:
Mass of scrap and crude iron
Steel grades specifications
Technical parameters of the oxygen-conversion & refining stages
Results of chemical analyses
Chemical composition requirements and standards for ferroalloy use
35
36. Process of building the ferroalloy absorption model
36
Attempt 1. Neural net. Failed.
Attempt 2. Straightforward regression. Element-by-element model.
Purely mathematical approach. Works, but seems to overlook some
important dependencies.
Attempt 3. Regression with MatrixNet (gradient boosting over
decision trees). Sophisticated feature engineering with the help of
domain experts. Success, but recommendations often look “odd” to
steelmakers.
Attempt 4. Model re-building to add more “physical sense” to it.
Machine learning used to “calibrate” traditional interpretable model.
Success! The concept confirmed.
Attempt 5. Experimentation with deep learning and other
techniques. Ongoing.
39. Smelting model. The details: three-steps modeling
Simple (e.g. linear) dependency on
the most important features 𝑧⃗ :
𝑧⃗ - Values of technical parameters
𝑦%
- Target (mass percent of
chemical element k)
𝑧⃗&, 𝑦%
- Historical dataset
𝑦%
≈ L(𝑧⃗)
More sophisticated dependency
on the whole set of features 𝑥⃗:
𝑦%
≈ F 𝑥⃗ =L 𝑧⃗ + M(𝑥⃗)
Probabilistic final model:
39
1 2 3
Smelting
model
Y D F
Probability
Amount of Mn
Permitted
chemical
range
L(z)
0 z 0
40. Optimization
40
The domain of confident
meeting the specifications
Threshold of confidence for
meeting the steel
specifications
Dopant2,kg
Dopant 1, kg
41. Timeline
41
Steelmaking process
Converter stage Refining stage Refining stage Refining stage Casting
The
furnace is
charged
Temperature and
oxidation measured
at the end of
blowing
Chemical
analysis
Chemical
analysis
I II III IV
Recommendations at the converter
stage
Recommendations
at the refining stage
42. Service step by step
1. Build a “classical” model that predicts the chemical
composition of steel; the model is based on the data
on historical smeltings
2. Evaluate the error of the model
3. Build the probabilistic model
4. Evaluate the domain of confident meeting the
specifications
5. Optimize the costs
6. Enjoy!
42
43. Human factor: acceptability of recommendations
43
Phase 1. Avoidance. (We couldn’t even know if recommendations
were good or not, because steelmakers didn’t want to follow them).
Phase 2. Incorporating the level of acceptance of recommendations
as a KPI for the service. (Led to re-building of the model).
Phase 3. Asking for comments on rejection reason.
Phase 4. Happy steelmakers!
51. How the solution is integrated
51
Resource intensive machine learning
happens in Yandex’s cloud
The resulting model is lightweight and
runs on premises
Solution has no own interface and is
integrated with existing internal
process management system
MMK
Model Training
Model transfer
Application
Evaluation
Data Upload
Prediction
YDF
53. Next steps and possible developments
Model extension to optimize the use of other ferroalloys
Model updates with new data to increase quality
Extension to other stages of production process, e.g. refining
Extension to other parameters of the process (e.g.
consumption of oxygen, fluxes etc.)
Other use cases (e.g. steel quality prediction).
53
54. Key takeaways
First time it always takes longer
〉Instead of 3 months it took 6 months twice
〉Model cannot be “slightly rebuilt”, it takes the whole cycle
〉Limited domain expertise is required, not only mathematics
It is crucial to have the right people on the ground
〉Integrating AI is not an IT task, it is a managerial effort
〉You will have to deal with new ways of assigning responsibilities
Choose the right pilot projects and work iteratively
〉You need clearly measurable economic effect
〉Focus on short-term PoC and not wide “roadmaps”
〉You will learn on the go54
!
55. Here’s our prediction
We know exactly what we want to
improve and can measure it
We have enough data
We can experiment
We can take automated action
In 10 years, we’ll have
algorithms doing the work
55
In all the business processes, where:
56. Will AI change the world in a decade?
Very likely.
Can it earn you a first million dollars this year?
Definitely.