30 |
Oilfield Technology
December
2014
less fanfare. Themethods used by soccer‑playing robots to recognise the
goal‑posts, other players and the ball are also used by Google’s self‑driving
car to understand the environment (and led to California passing the first
batch of laws for self‑driving vehicles on public roads this year), by Facebook
to recognise faces for tagging in pictures, and even in compact cameras
that have autofocus‑functions that automatically identify the human face.
Artificial intelligencemethods are found in the recommendations obtained
fromNetflix or Amazon, spam filters for email and routinely identify attempts
at credit card fraud. However, once themethods becomemature and
therefore ready for commercialisation, whatmattersmost is the product
or solution they are associatedwith rather thanwhatmethods are used
inside. Who cares if there is an artificial intelligence technology in the email
spam filter? The important thing is that there is no longer a deluge of emails
fromextremely rich, recently‑deceased, distant relatives or offers for herbal
remedies of dubious provenance clogging up the inbox. In fact, email spam
filters have become so good thatmany people have forgotten or will never be
aware that artificial intelligence is fundamental to themat all. It just works.
AI inthedigital oilfield
In oil and gas, this trend of artificial Intelligence technologies going
mainstream, not by performing feats of science fiction, but helping solve
business problems, can also be seen. There are hundreds of papers
published through the Society of PetroleumEngineers (SPE) and other
venues since the 1970s on the experimental use of artificial intelligence
techniques to solve problems big and small. This is typical also outside the oil
and gas industry, but in the last fewyears artificial intelligence techniques in
the formof predictive analytics have become important on a strategic level
to businesses inmany industries that have data, but also have challenges of
getting value fromthat data. In the oil and gas industry, this has been seen
particularly in the concept of the ‘digital oilfield’, where the idea is to collect,
store andmake available data fromthe field in centralised databases that
can support decision‑making and analysis around field operations. This
support can take the formof real time operation centres for specific services
such as geo‑steering, for general monitoring of production or a drilling
operation, or it can enable the rig crew to pull in expertise fromthe larger
organisation to examine the data for issues thatmay arise.
The first step in the upstream industry has been to collect andmake
available the data fromall forms of operations – drilling, casing, wireline,
fracturing and productionmonitoring. Then came efforts to seamlessly allow
the integration of this vast sea of data fromany vendor solutions through
standardised data transmission protocols such asWITSML and PRODML. This
mirrors a trend acrossmany industries in datawarehousing – the idea that
collecting bulk enterprise data allows newways of using that data.
The challenge in the next stage of the digital oilfield is to define
howexactly to use this data in new, innovativeways to drive efficiency
in the operations. For drilling operations, an experienced engineermay
be able tomonitor up to four operations at a time, but some operating
centresmonitor 30 rigs ormore. Staffing such a centre can become a very
expensive endeavour –with 24 hour shifts and extras for rotation coverage,
requirements can easily climb to one person per rig or higher. However,
inevitablymuch of the time for these engineers is spent simply being aware
of the situation in thewells they aremonitoring, not always solving problems
interacting directlywith the rig, especially for those organisations that
are not set up to provide targeted drilling optimisation. This is clearly an
economic problem, but it is also a problemof consistency and knowledge
management. Engineers are only human after all, and although theymay
know, for instance, howa slowdrop in pressure can be a sign of awash out,
itmay bemissed because they are tired, distracted, inexperienced, or simply
have a log scale set so that a small, gradual decrease is not visible.
This is the type of activitywhere computers excel – the patient,
consistent and unblinkingmonitoring of high volumes of data. However,
once a potential issue has been flagged, investigating it further and
addressing it can be a quite complicated process, possibly involving
time‑consuming detail analysis and physical modelling, considering the
overall situation of the operation and business goals. These creative problem
solving techniques arewhere people excel, andwhere engineers in a real
time operating centre can provide visible, valuable support to thewellsite.
In an example fromthe drilling processwheremachines provide the
consistency but humans are required for interpretation, when using a
common autodriller, its control loop logic can cause inefficient or damaging
states of fluctuation on the rotations perminute andweight‑on‑bit. This is
particularly visible if a real timemechanical‑specific energy (MSE) calculation
is available, and if examined at a relatively detailed resolution. These types
of fluctuations have proven to be very hard on downholemotors, causing
themto fail prematurely, which is a serious operational and supply chain
problem in some fields. This is an example of a pattern that is relatively
straightforward to use analytics or artificial intelligencemethods to spot.
Deploying such amethod to spot when an autodriller exhibits these
fluctuations takes away the need for engineers tomonitor for themmanually
and allows themto help in the resolution of the problemquicklywhere and
when it occurs instead.
Combining insight fromphysicsandanalytics
With the explosion of ‘Big Data,’ many commercial toolkits have included
methods fromartificial intelligence and statistics in advanced analytics
packages, and the hype is that these are nowso easy to use that it is simply
enough to point themat the data and insight and useful analysiswill come
out. For some data, this is in fact astonishingly close to the truth, as long as
there are people skilled in themethods to apply them. Google’s automated
service to translate between languages, for instance, has very little built‑in
knowledge of language or grammar – there is simply somuch text out there
already translated betweenmultiple languages that Google can use it as a
gigantic Rosetta Stone to learn how to translate.
In oil and gas, the complexities are not usually of language semantics,
rather that any reasonably intelligent analysis of data often requires deep
understanding of the physics of the process fromwhere the data is collected.
For instance, taking a standard neural network and training it on drilling
data in order to predict stuck pipe incidents is unfortunately bound to fail.
In theory, artificial intelligencemethods could even learn the basic physics
fromthe data, just like Google learns languages, but the amount of data
that would be required to do sowould be truly staggering. There are a lot of
wells drilled around theworld, but even that number is lowcompared to the
number of pages of text available toGoogle, or the number of pictures on
Facebook. Added to that is the fact that the incidents that need to be spotted,
such as a stuck pipe, are only a fraction of the total wells ever drilled. Methods
that require thousands ormillions of incidents to learn are simply not useful.
Currently the best solution to this is to combinemethods fromartificial
intelligence, advanced analytics and pattern recognitionwith knowledge of
the particular physics of, say, drilling awell, in a product targeted at solving
particular problems in the oil and gas industry. The good news is that this is
not science fiction or a research project any longer, but real tested products
available in themarket.
Theway to think aboutmodern, applied artificial intelligence and
analytics is less as a fantastically advanced (and high risk) approach, but
rather awell‑proven set of methodologies for certain types of engineering
problems. In particular, artificial intelligencemethods are good for narrow,
well‑defined problemswhere it seems it should be possible to automate
something people domore or less instinctively, but where it is difficult to
formulate an algorithm, rule, or exactmethod for doing so. For instance,
it is difficult towrite an algorithmto recognise a face in a picture, but it is
possible to use a pattern recognition technique to train it based on a set of
examples. Yet these engineering problems still need people skilled in the
methods and in the problemareas to build good solutions. The bottom line
is that although artificial intelligencewill always have a role in science fiction
dramas, it also has amuchmoremundane, but invaluable, role in day‑to‑day
life and business.
