Despite its faults and inaccuracies in early iterations, there’s no denying that AI is transforming our daily lives at an incredible pace and most of the time the features, and broadly speaking, the benefits it offers are extremely useful.
But in terms of its ability to completely transform the energy industry (and specifically oilfield) economics it’s important to consider why much of the early AI conversation needs to be tempered with a degree of objectivity.
The reality is one of marginal gains in many areas - much like creating a good sports team, over time these gains add up rather than causing instantaneous results everywhere. And this is especially the case within the oilfield technology industry:
As a short historical background on AI’s components, machine learning was introduced relatively early, when Frank Rosenblatt introduced the first artificial neural network (ANN) in 1958. Two years later Bernard Widrow and Marcian Hoff used this new technology to create MADELINE, an ANN that could eliminate echo in phone lines, which is still in use today.
Data Mining was born much more recently during the late 1980’s and it focuses on the discovery of previously unknown facts hidden in data. The simplest of these are correlations: For example, if two data streams are correlated, it can be assumed they are linked through a cause and effect relationship (although this isn’t always true nowadays).
In practice today
Some conversations in the energy industry suggest that AI can now transform everything in one giant leap. The fact is, as a highly valuable additional factor, it can and will provide significant incremental improvements. But it’s unlikely to radically transform everything overnight.
The successful cases in Oil and Gas will be those involving solid investments of engineering time and funding into the challenges faced. Decision makers also need to understand the difference and similarities in capabilities between the two specific areas within AI which are often confused - Machine Learning (ML) and Data Mining (DM). While similar theoretically, both employ the same methods and significantly overlap.
Alongside this consideration, the reality of AI in its current state is that it can be relatively complex to implement. In practice, large amounts of quality data are needed alongside time resource from very capable (and increasingly scarce) staffers to prepare that data. Characterising it appropriately and applying the different models can also make it expensive to implement in some cases.
Both Machine Learning and Data Mining do have huge potential to help us do things better, and examples already exist in practice within the oilfield arena. Again, it’s important to see these as "assistive technologies" in their current forms. Plugging them in to a data lake can’t solve all of the typical operational challenges we face – so we should also be wary of throwing out all the tools that currently serve their purpose effectively.
AI can already be easily used in combination with existing tools and first-principle models to enhance operational outcomes in our industry. It shouldn’t necessarily replace them, as many believe, and it’s worth noting that most successful examples of AI in effect today are in used in combination with other older technologies.
Real, actual and narrow AI
Artificial Intelligence is, by definition, not ‘real’ or ‘actual’ Intelligence. Just as artificial flowers from a distance look a lot like the real thing. Within this, ML is essentially about pattern recognition and it can be trained and refined to recognise patterns, so what we currently have is essentially "narrow AI". This involves carefully prepared data sets or trained ML models that specialise in solving a single particular problem. As a real-world example of this limitation, you can't take a model trained to drive a car and expect it to predict impending failure in a jet engine.
So, what is Narrow AI better at achieving than the human element? In our industry and among everyday applications it excels at processing large amounts of data and spotting such patterns. It’s also effective at identifying patterns earlier and more consistently than humans (for screening for cancer or searching for oil for example).
What Narrow AI is less effective at, however, is dealing with things it has never seen before and predicting things that have not occurred previously.
Recognising that a trained ML model is only applicable to the use case for which it has been trained is essential to the practical adoption of AI. It’s also important to understand that there are two basic types of ML tech – probabilistic models and classifiers.
Probabilistic models are trained on sequences of events, and can be interrogated to explain why a result was given. They can also be used to predict possible future events given an initial sequence.
Classifiers take in a set of information and make a prediction (or set of predictions) with associated probabilities; And neural Networks are good examples of these.
Considerations and concepts
A lot of the challenges we deal with in energy and other industrial settings are also time based - i.e. they are not discrete events, and they have momentum in the time dimension. The possible set of future states is dependent on the current state and the previous state(s). For example, if I am driving my car due north at 70mph now, in one second’s time the probability that I will be driving it south at 100mph is zero - there's no point in considering this outcome at all. I may have slowed a fraction, sped up a fraction or changed direction a little, and those possible changes can all be described by a set of (well known) equations.
Real world comparison
Perhaps the most well-known example of AI in use today is the self-driving vehicle. It’s a widely used reference for AI, and is therefore worth mentioning because it actually involves less AI than people think. It also depends heavily on existing techniques to provide Entity Recognition and Feature Extraction before the AI can come into play.
This is one where relevance recognition is easy – most people know how to drive!
Self-driving car technology is made up of a combination of image processing libraries (straight forward mathematical transforms of images) combined with algorithms that decide which transforms to apply to best identify the contents of an image (e.g. the lines on the road and other cars). A trained neural network (a classifier) then identifies the correct steering angle to apply, along with physical models of the car's performance, such as acceleration and braking, to keep within the speed limit and avoid collisions. Enormous amounts of training data is needed to be confident that the car will behave in the appropriate way.
Self-driving cars are a well-defined problem and it’s relatively inexpensive to go out and generate training data. You can also easily check the results. Industrial facilities are many times more complex, however. In practice, how do you figure out what the "lines in the road" are for a power station for example, when each item of equipment has its own ‘lines’. Generating test data is also challenging in our industry and checking the results requires complex visualisations and simulations.
So, to effectively use AI to solve a problem it needs to be a well-defined one with large amounts of training data available. It doesn’t help if you throw in “all” the data streams; It is more effective to pre-select the data that you really need. It also helps to be able to identify the relevant entities and have numerous example features. You also need to have people within your team that understand that data and be able to supervise the Machine Learning process.
Good data quality is also essential to support AI. The reality of many oil and gas facilities today is that the data quality is not good enough to support widespread implementation of AI unless remedial work is carried out on the data generating systems first. In some cases, this could be as simple as just increasing the frequency of data storage, while in others it will involve the addition or replacement of sensors, equipment and possibly work processes first.
So, what are the practical examples of Operational AI that can be implemented now, on existing facilities?
‘Risk reduction by identification of relevant previous incidents’ is a good one. I saw an excellent example at a recent oil industry event where a knowledge graph was built from an archive of over 100,000 incidents. This knowledge graph was used to identify likely risk levels for any planned activities based on the activity, equipment and location. Involving an activity that humans were having to do manually, it was well defined but it was also time consuming and open to errors. There was a huge amount of training data available and the output involved a recommendation to a human who then decides what outcome to take. This took a while to implement though - gathering and processing all the training data takes a long time. Classifying it is resource intensive and supervising and refining the training is also time intensive – but it can be done.
This example is highly relevant in an industrial context because the upsides are high – a large amount of time saved. And the downsides are low - you won't ‘blow anything up’ if the AI gets it wrong. The benefits are not "game changing" or industry disrupting but it is a very real improvement in a number of areas (particularly where safety and efficiency are concerned).
As with any mature industry, the improvements are often marginal but the key players keep making them. Those hoping to use AI to entirely transform the oilfield bottom line should be aware of its potential limitations - but also the clear opportunity for incremental improvement with the right degree of input.
Murray Callander, CEO, Eigen