Actionable Insights Towards

Competitiveness

© Aingura IIoT 2018

Smart Manufacturing: Data-based
actionable insights

Javier Diaz
IIoT Team Leader

© Aingura IIoT 2018

About Aingura IIoT

© Aingura IIoT 2018

 Aingura IIoT

Our objective is to use data and domain

knowledge to provide added value
bringing competitiveness to the industry,
at product and process level, through
machine learning-based failure diagnosis,
prognosis and energy efficiency
actionable insights.

© Aingura IIoT 2018 4

IIoT Use Case Example

© Aingura IIoT 2018

Using IIoT to increase machine availability

• Chinese automotive
OEMs
o More than 300 machines
working
• None of the above are
performing the same
operation
• However, all of them are
looking to:
o minimize downtime
o increase availability

© Aingura IIoT 2018 6

Using IIoT to increase machine availability
• Product
o Powertrain crankshaft
• Cycle time
o 60 seconds
• Average production
o 1.000 parts/day
• Required availability
o 95%
• Problem
o Large temperature gradients within
production facility
o Reduced availability when machine
stopped by low temperature
o Loss of precision
o Quality issues risk
o Machine stop could be up to 2
hours per day
o That is, more than 80 crankshafts
not produced.
o A stop machine can costs around
$50k per hour

© Aingura IIoT 2018 7

Using IIoT to increase machine availability
• Sampling rate
o Probe measurement: 240s
o Temperature: 80s
• Number of variables
o 15
• Main variables
o X and Y tooltip position,
o 9 machine structure and fluids
temperatures
o Environment temperature.
• Sampling time
o 12 months
• Total dataset size
o 2.4 GB
© Aingura IIoT 2018 8
Using IIoT to increase machine availability
• Machine Learning application
o Feature subset selection
 Select the most relevant variables (sensors) that has influence on the tooltip
position
o Multi-output regression
 Find how variables influence on the tooltip position
 Predict the tooltip position
 Provide feedback to the compensation control at the CNC
• Results:
o One part of the machine basement is the responsible for tooltip deviation
 New machine materials are studied for further design improvement.
o Compensation of the CNC system is improved by this model
• Outcome:
o To provide better knowledge from the machine to the designers
 Direct impact the machine design in terms of materials used and their specification
o Dynamical compensation of machine-tool behavior during production
o An increase crankshaft quality in terms of tolerance variation during
thermal changes and machine availability.
o An important increase in availability
 Avoiding machine-tool stop until stable environmental temperature is reached.
o Saved downtime costs up to $100k per day.
© Aingura IIoT 2018 9
 IIC Testbed:
Smart Factory Machine Learning
for Predictive Maintenance

© Aingura IIoT 2018

 SFML Testbed

http://www.iiconsortium.org/smart-factory-machine-learning.htm
© Aingura IIoT 2018 11
 SFML Testbed
• Sponsors:
o Aingura IIoT
o Xilinx
• Supporting:
o Aicas,
o Bosch Software
Innovations,
o GlobalSign,
o Infineon Technologies,
o iVeia,
o Microsoft,
• Phase 1: Lab Development and Test
o PFP Cybersecurity,
Utilizes simulated data and degradation/fault conditions for ML exploration
o RTI,
• Phase 2: Pilot Factory
o Thingswise,
Initial Deployment in limited production facility – Etxe-Tar
o Titanium Industrial Security,
and
• Phase 3: Production Facility
o XMPro
Deployment of ML and real-time analytics in Automotive OEM facility
© Aingura IIoT 2018 12
Preliminary Public Results

© Aingura IIoT 2018

Industrial Applications of Machine Learning
• Book details:
o Title: “Industrial Applications of Machine Learning”
o Series: Chapman & Hall/CRC Data Mining and Knowledge
Discovery Series
o ISBN 9780815356226 - CAT# K346412
o CRC URL: https://goo.gl/psf3Xi
o Table of Contents
1. The Fourth Industrial Revolution
2. Machine Learning
3. Applications of Machine Learning in Industrial Sectors
4. Component-Level Case Study: Remaining Useful Life of
Bearings
5. Machine-Level Case Study: Fingerprint of Industrial Motors
6. Production-Level Case Study: Automated Visual Inspection of a
Laser Process
7. Distribution-Level Case Study: Forecasting of Air Freight Delays
© Aingura IIoT 2018 14
Industrial Applications of Machine Learning

• Exploratory analysis
o Explore in the data without clear idea
o For small amounts of data, conventional
visualization methods
o For large amounts of data, dimensional reduction
• Example
o Real Application on machine tool
o Performance analysis of 3 servomotors
o 13 variables per servo
o 5 different algorithms:
 Agglomerative hierarchical clustering
 K-means clustering
 Spectral clustering
 Affinity propagation clustering
 Gaussian mixture model clustering

© Aingura IIoT 2018 15

Knowledge discovery with real data
• Testing 3 different clustering algorithms to find new
knowledge
o K-Means, agglomerative hierarchical, Gaussian mixture model.
o J. Diaz-Rozo, C. Bielza, and P. Larrañaga, “Machine learning-
based CPS for clustering high throughput machining cycle
conditions,” Procedia Manufacturing, vol. 10, pp. 997–1008, 2017.
• Machine-tool for powertrain manufacturing
o Cycle time 60 seconds
o Utilization over 95%
• Spindle head – Key critical component
o Power 10 kW
o Primary function: Material removal
• Failure cost :
o Costs USD 30,000 up to 250,000
o Repair time: 5 working shifts
o Impact: 200 direct jobs
• Understand Cluster Evolution:
o Cluster shapes (how the identified machining characteristics
change over time)
o Number of clusters (identify new machining characteristics).
• Gaussian mixtures
o Provides new information about different states of the spindle

© Aingura IIoT 2018 16

Gaussian-based Dynamic Probabilistic Clustering
• GDPC is an algorithm developed by
Aingura IIoT to measure component
degradation
o J. Diaz-Rozo, C. Bielza, and P. Larrañaga, “Clustering
of Data Streams with Dynamic Gaussian Mixture
Models. An IoT Application in Industrial Processes,”
IEEE Internet of Things Journal, 2018.
o https://doi.org/10.1109/JIOT.2018.2840129
• Data stream analytics
o Able to perform analytics in Real-Time
o No need of data storage
o Machine Learning at the edge
• Update the learnt model once the
component degrades
o Concept drift

© Aingura IIoT 2018 17

Edge Computing Node
• Integrated modules for:
o Analog sensors
o High speed energy
measurement
o Vibration
o Ethernet/switching
o Storage
• Powered by Xilinx MPSoC
Ultrascale+

© Aingura IIoT 2018 18

Edge computing deployment
• There could be a need for
computing power at the
edge
o Traditional computing devices
not suitable for industrial
environments
o Large amounts of data to be
pre-processed depending on
application
o Complex algorithms to solve
specific questions
o Extremely fast computing
needs to provide actionable
insights in Real-Time
• Steps for industrial
computing at the edge
US Patent 10031500B1
“Device and system including
multiple devices for supervision
and control of machines in
industrial installation”
© Aingura IIoT 2018 19
 Javier Diaz
IIoT Team Leader
jdiaz@ainguraiiot.com

Actionable Insights Towards

Competitiveness

© Aingura IIoT 2018