Russian oil producer goes digital; GE and Noble digitize their marine operations
Oil majors are getting wise to the possibilities of machine
learning, helping them manage risk more efficiently and with less subsea
inspection. That translates to the potential for substantial efficiency
gains and cost savings across the industry, Kieran Kavanagh, Innovation
and Data Analytics Director at Wood Group, said.
April, French oil major Total announced a tie-up with Google Cloud to
jointly develop AI solutions for subsurface data analysis in oil and gas
E&P efforts. Under the partnership, Total geoscientists will work
side-by-side with Google Cloud’s machine learning experts within the
same project team based in Google Cloud’s Advanced Solutions Lab in
The move reveals heavyweight backing for machine
learning in the oil and gas industry, with Total teams set to explore
multiple machine learning applications, including production profile
forecasting, automated analysis of satellite images, and analysis of
rock sample images.
Total has already used machine learning
algorithms to implement predictive maintenance for turbines, pumps and
compressors at its industrial facilities in 2013.
joint initiative with Google takes the partnership further, and alludes
to the growing interpenetration of AI businesses and big oil; Google
Cloud’s CEO Diane Greene has an oil and gas background, having worked
with oilfield services company Schlumberger on its DELFI software
environment, which leverages digital technologies including machine
learning to improve operational efficiency and deliver optimized
production at the lowest cost per barrel.
growth reflects the industry’s migration towards techniques that make
use of modern computing power’s ability to store and process sizeable
data sets. Best understood as a type of AI, machine learning uses
algorithms to draw conclusions through the study of these data sets.
industry figures are point out the benefits of such technologies. “The
thing about it is that typically it allows you to have the ability to
learn from experience (data); you look at what’s happened in the past,
and you learn from it,” Kieran Kavanagh, Innovation and Data Analytics
Director at Wood Group, an energy services company headquartered in
Aberdeen, told Upstream Intelligence.
Supervized and unsupervized
are two main types of machine learning – supervized and unsupervized.
Supervized is when you know what “bad” looks like; that is there is a
sufficient “training set” to tell a computer how to recognize patterns
“Say, for example, you have experienced some
degradation or failure of a piece of equipment or component in the past.
You can then look at the history of monitored data relating to that
component (for example temperature or vibration information), to see if
the measured data can be related to the onset of degradation,” says
Put simply, if you know what poor behaviour looks like,
you can go and identify that and teach the machine – introducing an
algorithm to recognize data patterns that you have seen before.
cited an example: if there's a time trace of measured temperature or
pressure at a well, and a certain behaviour patterns in this data is
observed, this could be indicative of the start of degradation or
failure. “You can now say that each time you see that behaviour again,
it may be indicative of degradation or failure. With a good history of
measured data, linked to known (tagged) events, we can use this
‘training’ data from the past to predict the future.”
Caption: Valve signature machine learning – identifying characteristics of healthy or degrading vales.
Source: Wood Group
Unsupervized machine learning is when it may not be known what
degradation or failure look like, but there is a view as to what
non-normal looks like.
Technology providers are now actively
marketing machine learning products tailored for the oil and gas
industry. Switzerland-based Akselos, a digital twin developer, it is
attempting empower predictive maintenance with machine learning
Thomas Leurent, Akselos’ Chief Executive told
Upstream Intelligence that Akselos an active R&D programme to
enhance machine learning, and make it part of the technology ecosystem.
use a highly accurate and operationalized physics-based model which is
basically about augmenting predictive power; that’s why we are doing
it. Machine learning is mostly about predictive power -- sometimes it
is about learning from failure, but mostly it is about predictive
power,” he said.
the work that goes into building predictive power at the design phase,
but that has tended to be lost the moment we go into operations. “That
has to stop,” said Leurent. “We need that predictive power in operations
now, as part of the digital transformation. “
Akselos have taken
the technology and made it operations-ready, speeding it up by factor of
1,000 and sometimes a lot more. And it has coupled it with sensors that
learn from sensors that are deployed in operations – and that has made
the workflow much easier, so it can be used in operations.
Caption: Akselos created the world’s largest predictive digital twin of a 90,000 ton FPSO.
“Essentially, we’ve changed predictive technology to make it
operations ready,” said Leurent. “We try to couple machine learning and a
subset of machine learning that is uncertainty quantification. In the
upstream industry, there is plenty of uncertainty of the inputs. It is
all very good to be using model, but if you don’t know what the input
is, then we have a problem.”
Machine learning has enormous
potential in the integrity management side of the oil and gas business.
Wood Group’s Kavanagh highlighted the implementation of a machine
learning solution is in what it calls “Valve Signature Analytics”.
addresses the challenge the subsea industry faces in identifying when
valves might be degrading. Wood has applied machine learning analytics
to the time trace of key operating parameters that are monitored for
these valves. In this case, said Kavanagh, it managed to implement a
solution that has allowed it to anticipate degrading and failure
typically 1 to 2 years before the failure occurs.
approaches can be used to reduce the frequency and costs of subsea
inspection. “For inspection costs that could run to $100,000 a day for
lengthy programmes that may be carried out each year, the potential cost
reductions in this area are clear,” said Kavanagh. “If machine learning
helps you manage risk more efficiently and with less subsea inspection,
that translates to the potential for substantial efficiency gains and
cost savings across the industry.”