S8714

Deploying Machine Learning on the

Oilfield: From the Labs to the Edge.
- Matthieu Boujonnier – Analytics Application Architect
- Bartosz Boguslawski – Data Scientist
- Loryne Bissuel-Beauvais – Data Scientist

Confidential Property of Schneider Electric

 Are you ready to
deploy your data scientists’
work on this pump?
 Are you ready to
deploy your data scientists’
work on this pump?
Yes! Those guys!!!
 Are you ready to
deploy your data scientists’
work on this pump?
And that works also for
Yes! Those guys!!!

…cars

…evil robots
Our Challenges

ENSURE TRUST
How can our customers trust ML
predictions?

EXTEND MODELS
Labeled « expert » data is rare, how to ensure that
our models will work for any « new » pump?

Confidential Property of Schneider Electric | Page 13

Realift: A Pilot on Rod Pumps
Confidential Property of Schneider Electric | Page 14
The Digital Oilfield Anatomy
Past, Present and Future

PRESENT

AUTOMATION Fewer and fewer

expertise available:

Local workforce needs to be

empowered
DATA FUTURE

OPTIMIZATION

EXPERIENCE

Confidential Property of Schneider Electric | Page 15

Rod pump: facts

38% of the total production

750 000 Wells Impact of downtime at
CURRENT
GLOBAL OIL CURRENT DAILY
$65 per barrel of oil
PRODUCTION SITUATION PRODUCTION

80% of US
74 M RPC make 5-2.5 K
BPD BPD
less than 10
BPD

AVERAGE RUN LIFE OF A RPC OPTIMAL RUN LIFE OF A RPC

3 years 10 years

Confidential Property of Schneider Electric | Page 16

Today’s SCADA solution

Confidential Property of Schneider Electric | Page 17

Customer’s confidence is key
ML
Predictions

Insights /
Advice

Field
Services

Stop production /
Change equipment

Generate
costs

Confidential Property of Schneider Electric | Page 18 © XKCD

Customer’s confidence is key

REALLY?

Confidential Property of Schneider Electric | Page 19 © XKCD

Customer’s confidence is key

Let’s do like humans do !

Confidential Property of Schneider Electric | Page 20 © XKCD

 A bit of mechanics and mathematics

Gibbs’s Wave Equation

Miles below
the surface

Source: http://petrowiki.org,
Schneider Electric

Confidential Property of Schneider Electric | Page 21

Easier with a little animation…

Confidential Property of Schneider Electric | Page 22

Human expertise
• Experts look at charts of failure patterns
• Experts use mostly their eyes, which interpret the image to a failure (but also look at the data)

Expert

Simplified Dynacards showing a failure pattern

Confidential Property of Schneider Electric | Page 23
The way Machine Learning usually works

1) Training: build model

Training

data Model label Cat

2) Inference: use model

Inference
Cat (0.86)
data Model label
You (0.14)
(after this session)

Confidential Property of Schneider Electric | Page 24

The way Machine Learning usually works

1) Training
It’s very
Data challenging to
Label
Model obtain very large
Data Label amount
of labeled data
2) Inference

data Model Label

Confidential Property of Schneider Electric | Page 25

Data augmentation
Using existing tagged cards to increase the size of the training dataset
One way to get around a lack of data is to augment the dataset. The model will often be more robust and
can even be simpler due to a better training set. This may prevent overfitting as well.
Pick images of the same class and combine them to get a new one:
gas lock 1 gas lock 2 new gas lock

+ =

worn pump 1 worn pump 2 new worn pump

+ =

Confidential Property of Schneider Electric | Page 26

Dynacards can be considered as an image, right?
Convolutional Neural Networks
CNN typical architecture

Model: “Those parts of the dynacard Model: “Mostly because of those parts of
indicate mostly to me that this pump is the dynacard I think this pump works
grinding” perfectly well”
Confidential Property of Schneider Electric | Page 27
Not sure what kind of cat it is but for sure not a dog!
Siamese Network

When humans look at an object, they recognize its class not only because it’s similar to some object
but also because it’s different from some objects.

Image 1: gas lock

Conv. Neural Net

Binary Output
1: same
0: different
Conv. Neural Net
Image 2: pump grinding dense layer
Labeled at learning time concatenation
Unlabeled at inference time

• Training is done by comparing each image against all the images in dataset and checking if the
class is the same or different
• Data set is “augmented” by the combinations of pairs of all the images available

Confidential Property of Schneider Electric | Page 28

Not enough labeled data? Use autoencoders!

Reconstructed
Input image Latent space image
representation

encoder decoder

Self-supervised model - trained without labels!

Fully-connected network
Latent space
representation

Confidential Property of Schneider Electric | Page 29

Dynacards are shapes, right?
Histogram of Oriented Gradients (HOG)

Simplify image by extracting new features:

Gradients (x and y derivatives) of an image are large around contours (regions of abrupt intensity
changes) and we know that contours contain a lot more information about shape than flat regions
especially in our application!
Reduced dimension
Extracted features

Model Label

Confidential Property of Schneider Electric | Page 30

Increase your odds! Use ensemble of models!
Instead of having one model combine many of them and make our task a team work!

Input data Run all models Final output

Weights
CNN

pluid pound
gas interference
Siamese
solids grinding
Ensemble
gas lock
model normal
AAE+FCN
plunger stuck
0 0.5 1

HOG

Confidential Property of Schneider Electric | Page 31

Closing the loop: the Edge deployment
Confidential Property of Schneider Electric | Page 32
From the labs to a solution ready to sell
Data Acquisition Data Exploration & Detection Results
Data is collected from Dynacard Pattern recognition Dynacard Pattern evolution RPC immediate
local systems diagnostics

Reduce
safety
risks

Reduce
downtime
RPC Data Cleaning /
Segmenting
Increase
Data Preprocessing Production

Reduce
maintenance
costs

Confidential Property of Schneider Electric | Page 33

 Realift Architecture, the marketing speech

The solution is deployed in

harsh environments where

• Internet connectivity is
unreliable and expensive
• Low bandwidth
• Customers require high
data privacy and
confidentiality
• Critical systems are
installed

As a result, Realift® is a full

Edge solution that does not
rely on “always available”
connectivity!

Onsite deployment

Confidential Property of Schneider Electric | Page 34

Enhanced with Azure IoT Edge

GPRS not always-on

connectivity

Confidential Property of Schneider Electric | Page 35

Use transfer learning to adapt the models locally
Local screen Feature Extractor Classifier

Training
data

Transfer

Transfer Learned knowledge

“Feedback” labelled
Local dataset Back propagation

Confidential Property of Schneider Electric | Page 36 Frozen weights

 Questions and Answers
Confidential Property of Schneider Electric | Page 37
Resources
• Deep dive on the ML model:
https://tinyurl.com/devintersect
• WSJ article:
https://tinyurl.com/wsjpump
• Microsoft Customer Story:
https://tinyurl.com/mscuststory Learn More

Confidential Property of Schneider Electric | Page 38

Realift™ installation in North Dakota