choosing your first ai project. how to get a quick roi in process industries
TRANSCRIPT
Choosing your first AI project: How to get a quick ROI
│ 28 November 2017
Current state of the industryOverwhelmed with digitalisation
Smart factory
How to close the gap?
And how to fund all these innovations?
Capital investments
Process redesign
Lengthy deployment
ROI in 5-10 years
And how to fund all these innovations?
Capital investments
Process redesign
Lengthy deployment
ROI in 5-10 years
… Results are not guaranteed!
Agenda
1. What industrial AI is
2. How to map specific processes fit for AI
3. How to prioritise AI projects
4. Bonus: what is wrong with predictive maintenance
+ Q&A
What is AI and machine learning
Complex algorithms that:
〉accomplish tasks by themselves instead of being explicitly programmed
〉learn inductively from past historical data (or data generated for training)
〉can predict events or prescribe actions
What is AI and machine learning for industrial sector
Newbusiness
models
New products orservices
Process automation and optimisation
The ”invisible” AI
Low capital investment
Quick wins, ROI < 1 year
How it differs from systems already in use?
Processes relying on traditional knowledge-based models
Results of chemical analyses
Equipment telemetry
Process parameters
Traditional models of physical processes
embedded in process control
systems
Expert judgement
L(z)
0 z
Processes relying on traditional knowledge-based models
Results of chemical analyses
Equipment telemetry
Process parameters
L(z)
0 z
How AI creates value
Learns from how the process actually ran on specific plant
Accounts for many weak dependencies in past data
Operates on top of knowledge-based models
“Personalises” production decisions in every iteration
New level of precision for the cases where it really matters
Complementary to existing process control
Direct effect with no capital investment
Steps to choose your AI project
Step 1: Start with business needsEstablish the foundation
Start with business needs
“Let’s do deep learning / chatbots”-> Technology for its own sake
“Let’s buy sensors / clean up databases”-> Data generates costs, not value-> Postponing outcomes for no reason
–
Start with business needs
Review the usual pain points
Quality control and assurance
Productivity and yield
Energy consumption
Raw material use
+
Start with business needs
High-volume, low-margin product-> Decrease raw material use
Recurring fluctuations in quality-> Decrease losses
Specific shop is a bottleneck-> Improve throughput
+
Step 2: Identify where AI is applicableScan the use case horizon
Uncertainty and complexity
Process has fluctuations
Many factors involved
Some parameters not known with precision
Non-linear dependencies
Often experienced operator makes the decisions
You need advanced process control
? ??
??
Uncertainty and complexity: gas fractionation
Sources of complexity:
— Continuous process
— Chemical composition of gas feed varies
— Adjusting parameters (e.g. temperature) takes time
— Changing them too fast may lead to disruptions
Optimisation goals:
— Improve throughput / energy efficiency
Uncertainty and complexity: examples
Oil and gas well drilling(increase drilling speed,
avoid deviations)
Animal food production(optimise moisture content,
decrease variability)
Blast furnace process in steelmaking (decrease
energy use)
Uncertainty and complexity
Perfect “lab” environment
Low number of measured / controlled factors
Rule-based systems work well enough
–
Established, repetitive process
Historical data is available
Historical data remains relevant to rely for future
Even small improvements are significant due to process frequency
Optimisation makes long-term business sense
The process is likely already efficient enough
Established processes: steelmaking
— Equipment lifespan is typically decades
— Core processes are essentially same
Possible applications
— Decrease raw material (ferroalloy, fluxes) use in oxygen converter
— Decrease energy use in electric arc furnace
— Decrease defect occurrence during rolling
Established process: examples
Catalytic cracking Beer fermentation Bottle counting
Established process
One-time, irregular, or too diverse processes with insufficient historical data
〉Design a new plant
〉Invent new product recipe
〉Forecast market trends
–
Measurable KPI
AI has no ”common sense”, it needs a metric to optimise
Should be as close to business as possible
Should be measured in a straightforward way
Measurable KPI: optimisation of ferroalloy use
— Goal: Decrease the use of raw material without affecting steel quality
— Metric: Average costs of ferroalloys in smeltingsperformed following AI recommendations
— Restriction: Chemical composition of steel should fit in required ranges (specification)
5%average decrease
>$4.3myearly effect
Measurable KPI: optimisation of ferroalloy use in steelmaking
$$$$$$Optimisation potential
$$$Cost savings achieved
KPI: optimisation of raw material use
Zinc use in steel coating
Cyanide use in ore leaching
Cocoa butter use in chocolate conching
Summing up: prerequisites for AI
Established,repetitive process
to be able to rely on data
Summing up: prerequisites for AI
Established,repetitive process
Uncertainty and complexity
for AI to create value to be able to rely on data
? ??
??
Summing up: prerequisites for AI
Established,repetitive process
Uncertainty and complexity
Well-defined, measurable outcomes
for AI to create value to be able to rely on data to measure success
? ??
??
Step 3: Prioritise the ideasReaching the low-hanging fruits
3 criteria
Data availability Expected effect Time to value
Data availability
Not necessary “big”
Starting from 10000+ process iterations, 1-2 years of logs (could be less for frequent processes)
Raw data
Errors and gaps are not that crucial
+
Data availability
“We measure it, but do not store”“We store only the last month, and then delete”-> Algorithm will have nothing to learn from
Significant process changes in history (e.g. full revamping of the line)-> Makes past data obsolete
–
Expected effect
Alternatives
Direct results are significant for business
Limited pilot that can still confirm the case
Conscious choice to do R&D to learn from
+
Expected effect
20-40% of total costs spent on cyanide in ore processing
Even 1% decrease in ferroalloy use is business significant
30% of total costs on testing costs and yield losses in semiconductors
Expected effect
Direct results are too small
〉Total quality losses on the line are only $50K per year
Absence of specific quantifiable metric
〉“We want to discover route causes of defects”
Inability to influence the outcome
〉Quality predicted at point when nothing can be done
–
Time to value
Scope of the project involves only one production stage and team
Data is readily available or is extracted easily
Experimentation easily possible
+
Use case definition
Data gathering
Model training
Model testing
Production use
Time to value: slab quality prediction
— Goal: Identify slabs that are likely to lead to defects during rolling
— Metric: Portion of slabs with high defect mass revealed
— Data: Historical data on 17000 slabs
— Pilot scope: Quality forecast tested historically.
— Outcome: Total pilot length less than 3 months. Ability to establish business case given known processing costs.
Time to value: quality prediction
Metals Plastics Glass and optical fiber
What is wrong with predictive maintenance
+
Definitely huge potential economic effect
–
Either too few examples of failures or minimal economic effect
Lots of noise and anomalies
Difficult to measure direct economic effect in short term
Predictive maintenance: approaches for pilots
Change the focus
Predict defects that occur due to equipment wear-out
Specify the object
Choose specific equipment element with frequent failures
Shift the task
Predict demand for spare parts
Predicting anomalies
Multi-stage process, combination of AI and human expertise
Summing up
Why you should use artificial intelligence for process optimisation
No capital investments
No disruption of existing process
3-6 months to implement
Immediate ROI
Capital investments
Process redesign
Lengthy deployment
ROI in 5-10 years
Why it is important to start now?
When best practices are established, it is already too late
AI is too different from traditional software products, and many things are to be learned from experience
Organisation and management will be affected, and it’s important to learn from experience to be able to prepare
Q&A
Elena SamuylovaMarketing and business development director
Emeli DralChief data scientist