s8714 deploying machine learning on the oilfield: from the labs … · 2018. 3. 30. · dynacard...

Deploying Machine Learning on the

Oilfield: From the Labs to the Edge.

S8714

Confidential Property of Schneider Electric

- Matthieu Boujonnier – Analytics Application Architect

- Bartosz Boguslawski – Data Scientist

- Loryne Bissuel-Beauvais – Data Scientist

Are you ready to

deploy your data scientists’

work on this pump?

Are you ready to


work on this pump?

Yes! Those guys!!!

Are you ready to


work on this pump?

Yes! Those guys!!!And that works also for

…cars

…evil robots

ENSURE TRUST

How can our customers trust ML

predictions?

EXTEND MODELSLabeled « expert » data is rare, how to ensure that

our models will work for any « new » pump?

Page 13Confidential Property of Schneider Electric |

Our Challenges

Confidential Property of Schneider Electric |

Realift: A Pilot on Rod Pumps


The Digital Oilfield AnatomyPast, Present and Future

Fewer and fewer

expertise available:

Local workforce needs to be

empowered

EXPERIENCE

OP

TIM

IZA

TIO

N

AUTOMATION

PRESENT

FUTUREDATA


Rod pump: facts

3 years 10 yearsAVERAGE RUN LIFE OF A RPC OPTIMAL RUN LIFE OF A RPC

74 MBPD

80% of US

RPC make

less than 10BPD

5-2.5 KBPD

CURRENT

GLOBAL OIL

PRODUCTION

DAILY

PRODUCTION

38% of the total production

750 000 Wells

CURRENT

SITUATION

Impact of downtime at

$65 per barrel of oil


Today’s SCADA solution


Customer’s confidence is key

© XKCD

ML Predictions

Insights / Advice

Field Services

Stop production / Change equipment

Generate costs



© XKCD

REALLY?



© XKCD

Let’s do like humans do !


A bit of mechanics and mathematics

Gibbs’s Wave Equation

Source: http://petrowiki.org,

Schneider Electric

Mile

s b

elo

w

the

su

rface

http://petrowiki.org/


Easier with a little animation…

• 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)


Human expertise

Expert

Simplified Dynacards showing a failure pattern


1) Training: build model

2) Inference: use model

The way Machine Learning usually works

Modeldata label

Modeldata labelCat (0.86)

You (0.14)

Cat

(after this session)

Training

Inference


1) Training

2) Inference

The way Machine Learning usually works

It’s very

challenging to

obtain very large

amount

of labeled data

Model

Modeldata Label

Data

Data

Label

Label


Data augmentationUsing 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


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

CNN typical architecture

Model: “Those parts of the dynacard

indicate mostly to me that this pump is

grinding”

Model: “Mostly because of those parts of

the dynacard I think this pump works

perfectly well”


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.

Not sure what kind of cat it is but for sure not a dog!

Conv. Neural Net

Conv. Neural Net

dense layer

concatenation

Binary Output

1: same

0: different

Image 1: gas lock

Image 2: pump grinding

Labeled at learning time

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

Siamese Network


Not enough labeled data? Use autoencoders!

Input image Latent space

representation

Reconstructed

image

encoder decoder

Self-supervised model - trained without labels!

Latent space

representation

Fully-connected network


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!

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

Reduced dimension

Extracted features

Model Label


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

CNN

Siamese

AAE+FCN

Ensemble

model

Input data Run all models Final output

0 0.5 1

plunger stuck

normal

gas lock

solids grinding

gas interference

pluid pound

Weights

HOG

Closing the loop: the Edge deployment


From the labs to a solution ready to sell

Data Preprocessing

Data is collected from

local systemsDynacard Pattern recognition

Reduce

downtime

RPC immediate

diagnostics

Data Acquisition Data Exploration & Detection Results

Data Cleaning /

SegmentingRPC

Reduce

safety

risks

Increase

Production

Reduce

maintenance

costs

Dynacard Pattern evolution

Confidential Property of Schneider Electric | Page 33


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

Realift Architecture, the marketing speech


Enhanced with Azure IoT Edge

GPRS not always-on

connectivity


Use transfer learning to adapt the models locally

“Feedback” labelled

Local dataset

Training

data

Frozen weights

Classifier

Transfer Learned knowledge

Feature Extractor

Transfer

Back propagation

Local screen


Questions and Answers

• Deep dive on the ML model:

https://tinyurl.com/devintersect

• WSJ article:

https://tinyurl.com/wsjpump

• Microsoft Customer Story:

https://tinyurl.com/mscuststory

Resources


Learn More

Realift™ installation in North Dakota

https://tinyurl.com/devintersect

https://tinyurl.com/wsjpump

https://tinyurl.com/mscuststory

s8714 deploying machine learning on the oilfield: from the labs … · 2018. 3. 30. · dynacard...

Documents