dc91439: gpu accelerated iiot data science and … · bryan massie and tony demarco 5 –november...

27
Bryan Massie and Tony DeMarco 5 – November – 2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE DATA LAKE

Upload: others

Post on 13-Jul-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

Bryan Massie and Tony DeMarco

5 – November – 2019

DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE DATA LAKE

Page 2: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

2© 2019 Lockheed Martin Corporation. All rights reserved.

SPEAKERS

Bryan MassieLM Fellow Enterprise IT

Tony DeMarcoPrincipal Data Scientist AERO Enterprise Integration

Page 3: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

3© 2019 Lockheed Martin Corporation. All rights reserved.

OBJECTIVES

• Illustrate IIoT function at Lockheed Martin Aeronautics

• Demo analytics model delivering business value

AGENDA• Operational Reference Architecture

• Data Pipeline

• Analytics Approach

• Anomaly Detection Use Case

Page 4: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

© 2019 Lockheed Martin Corporation. All rights reserved. 4

IIOT IS A KEY ENABLER FOR THE FACTORY OF THE FUTURE

Page 5: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

© 2019 Lockheed Martin Corporation. All rights reserved. 5

Operational Reference ArchitectureWing Drills

Page 6: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

© 2019 Lockheed Martin Corporation. All rights reserved. 6

Operational Reference ArchitectureWing Drills Autoclaves

Page 7: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

© 2019 Lockheed Martin Corporation. All rights reserved. 7

Operational Reference ArchitectureWing Drills

Paint Robots

Autoclaves

Page 8: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

© 2019 Lockheed Martin Corporation. All rights reserved. 8

Operational Reference ArchitectureWing Drills

Paint Robots

Autoclaves

Adapter

Allen BradleyFanuc

Siemens 840D

Asset Controllers

Page 9: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

© 2019 Lockheed Martin Corporation. All rights reserved. 9

Operational Reference ArchitectureWing Drills

Paint Robots

Autoclaves

Adapter

Allen BradleyFanuc

Siemens 840D

Asset Controllers

Cisco Industrial Ethernet 4000 Series

Mazak Smart Box

Page 10: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

© 2019 Lockheed Martin Corporation. All rights reserved. 10

Data Pipeline

Secure, Low-latency IIoT Data from Shop Floor to Top Floor

Page 11: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

11© 2019 Lockheed Martin Corporation. All rights reserved.

IIOT ANALYTICS APPROACH

Initial approach pursued during 2018

Drawbacks• Modeling requires significant

work by a knowledgeable human – knowledge of both the machine and the modeling approach

• A model for one machine class is not transferable to another i.e., not scalable)

Physics Based Modeling Machine Learning Physics Based Model Enhancement

Long Short-Term Memory Recurrent Neural Network

Upside✓Machine class agnostic

✓ Faster to produce meaningful results and identify anomalies / predict failure

✓ Scalable across machine classes

Augment Machine Learning approach with traditional physics-

based modeling

Pre-Estimate Fusion• Feature engineering• Data Selection

Post-Estimate Fusion• Ensembling of models

Traditional Approach Cognitive Approach

Page 12: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

12© 2019 Lockheed Martin Corporation. All rights reserved.

REFERENCE

• P. Malhotra et al. Long Short Term Memory Networks for Anomaly Detection in Time Series. Proceedings of the European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning

• https://www.elen.ucl.ac.be/Proceedings/esann/esannpdf/es2015-56.pdf

Page 13: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

13© 2019 Lockheed Martin Corporation. All rights reserved.

LSTM RECURRENT NEURAL NETWORK

• RNNs “remember” past observations; allows for using a sequence as in time) as input

• LSTM is an improvement over “vanilla” RNN

• Overcomes vanishing/exploding gradient problem

Page 14: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

14© 2019 Lockheed Martin Corporation. All rights reserved.

LSTM RECURRENT NEURAL NETWORK

• RNNs “remember” past observations; allows for using a sequence as in time) as input

• LSTM is an improvement over “vanilla” RNN

• Overcomes vanishing/exploding gradient problem

ℎ𝑡−1

𝑥𝑡

𝜎ℎ𝑡

Page 15: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

15© 2019 Lockheed Martin Corporation. All rights reserved.

LSTM RECURRENT NEURAL NETWORK

• RNNs “remember” past observations; allows for using a sequence as in time) as input

• LSTM is an improvement over “vanilla” RNN

• Overcomes vanishing/exploding gradient problem

Cell state “memory”)

Page 16: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

16© 2019 Lockheed Martin Corporation. All rights reserved.

LSTM RECURRENT NEURAL NETWORK

• RNNs “remember” past observations; allows for using a sequence as in time) as input

• LSTM is an improvement over “vanilla” RNN

• Overcomes vanishing/exploding gradient problem

forget gate

Page 17: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

17© 2019 Lockheed Martin Corporation. All rights reserved.

LSTM RECURRENT NEURAL NETWORK

• RNNs “remember” past observations; allows for using a sequence as in time) as input

• LSTM is an improvement over “vanilla” RNN

• Overcomes vanishing/exploding gradient probleminput gate

Page 18: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

18© 2019 Lockheed Martin Corporation. All rights reserved.

LSTM RECURRENT NEURAL NETWORK

• RNNs “remember” past observations; allows for using a sequence as in time) as input

• LSTM is an improvement over “vanilla” RNN

• Overcomes vanishing/exploding gradient problemoutput gate

Page 19: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

19© 2019 Lockheed Martin Corporation. All rights reserved.

MACHINE LEARNING FOR ANOMALY DETECTION

• A common approach to anomaly detection is to compare a predicted state to a measured state• If the measured state differs from the predicted state, something may be abnormal• The predicted state comes from some physics-based or statistical model• Here, the network is learning the operation of the production machine from historical data

Page 20: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

20© 2019 Lockheed Martin Corporation. All rights reserved.

ANOMALY QUANTIFICATION USING ERROR DISTRIBUTION

• One variable case is simple, we just look at the magnitude of difference between predicted and observed the error)• For more than one variable, it is far more informative to look at the errors adjusted for the covariance between them

𝒆𝑛𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 = 𝒆 − ത𝒆 𝐕𝚲−𝟏𝟐𝑽−𝟏

original

centered

align axes w/eigenvectors

scale by eigenvaluesrestore original orientation

Page 21: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

21© 2019 Lockheed Martin Corporation. All rights reserved.

ANOMALY QUANTIFICATION USING ERROR DISTRIBUTION

• One variable case is simple, we just look at the magnitude of difference between predicted and observed the error)• For more than one variable, it is far more informative to look at the errors adjusted for the covariance between them

𝒆𝑛𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 = 𝒆 − ത𝒆 𝐕𝚲−𝟏𝟐𝑽−𝟏

original

centered

align axes w/eigenvectors

scale by eigenvaluesrestore original orientation

Page 22: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

22© 2019 Lockheed Martin Corporation. All rights reserved.

IMPLEMENTATIONDomino Data LabR StudioNVIDIAKeraTensorFlow

Page 23: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

23© 2019 Lockheed Martin Corporation. All rights reserved.

EXAMPLE RESULTS

Page 24: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

24© 2019 Lockheed Martin Corporation. All rights reserved.

EXAMPLE RESULTS

Page 25: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

25© 2019 Lockheed Martin Corporation. All rights reserved.

Takeaways

• IIoT is a key component/driver of the Future Factory

• IIoT data pipeline enables secure, scalable, rapid value delivery

• Machine Learning models add value by detecting anomalies; helping to avoid costly unplanned maintenance

Page 26: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

26© 2019 Lockheed Martin Corporation. All rights reserved.

Contact Us!

• Massie, Bryan S (US) [email protected]

• DeMarco, Antonio (US) [email protected]

Page 27: DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND … · Bryan Massie and Tony DeMarco 5 –November –2019 DC91439: GPU ACCELERATED IIOT DATA SCIENCE AND MACHINE LEARNING IN AN ENTERPRISE

27© 2019 Lockheed Martin Corporation. All rights reserved.