Kicking off this mini-series last month, I promised that this time around I would talk more about the kinds of machine learning (ML) available to us and how they might be applicable in the world of engineering and heavy assets.
Just to set the scene, you may remember that last month I defined ML as the ability of computers to discover insights without being explicitly told where to find them1. Bearing that in mind, there are a few ways we can give computers the start they need to be able to come up with these new insights. That’s what I’ll discuss here.
Possibly the most intuitively obvious machine learning technique and therefore the one most people would be most comfortable with is supervised learning. It’s a lot like us learning from a teacher, who provides some example inputs and outputs, then says “off you go and find some more cases like that”. The computer learns to define a general case from the specific ones the teacher (ie us) has shown. Teaching continues until the teacher is satisfied that the results are sufficiently accurate.
Now, let’s think of applying supervised learning techniques in asset management. Our computer will be able to learn from training on historical data to categorise assets and understand paths that lead to failure, or maloperation or deviations from acceptable parameters or…whatever we have decided is important. And our computer will also be trained by example to react in appropriate ways when it sees those indicators in new data. It might raise an alarm. Or perhaps our computer will be trained to adjust certain control parameters that bring operations back on track. It may also regularly report detailed findings to assist in planning longer term changes to maintenance practices. It will, in fact, take any manner of action or intervention we have taught it by example to carry out.
As the name suggests, in this case there is no teacher. Though our computer is already programmed to know how to apply multiple generic analytical techniques to any data it is presented with. Next, it will be given unlabelled data and asked to categorise what it finds by applying those analytical techniques without being directed as to what is “right” or even what is interesting. Any correlations, segmentations or categorisations the computer then comes up with will be real (assuming the data provided is an accurate representation of the situation), but they may or may not be valuable.
Unsupervised learning is often used when it is too difficult or too expensive a problem to create training data sets or when a discovery approach that may lead to insights we might never have come up with ourselves2 is desirable.
Once we understand and ascribe value to them, these new insights might be an end in themselves. But they can also become inputs to supervised learning cases. For example, unsupervised learning may find an apparently unrelated group of assets that behave entirely differently from the rest of the fleet. Our ML computer won’t know why these assets behave in this way. It’s still our job to investigate further and work that out. But it will show us how they behave and how those behaviours are different to other groups of assets. If this information is now added to supervised learning training data, it may contribute to making even more accurate behaviour predictions than before.
Our earlier example of adjusting parameters in an operational system is a good one to discuss further when considering reinforcement learning. This technique uses trial and error to determine which set of actions leads to the greatest reward. Yes, our computer can already be trained to adjust control parameters that do indeed bring operations back on track. But reinforcement learning will train it to do so in the quickest time; or with the least process disruption; or with the right balance of risk and cost, depending on what we decide are the goals to be achieved. This is done by giving feedback about the outcomes of actions our computer takes. Actions that achieve the goals we have provided: good. Actions that don’t, or lead to disruption or increased risk or cost: bad. In this way, the computer can determine the best strategy for delivering the desired goals, without us telling it what that strategy is. Because let’s face it – we’d probably be wrong.
Yes David, but is it going to take over the world?
As I’ve tried to make clear above, the basic concepts of ML are fairly simple and eminently understandable. When we look behind the hyperbole and the marketing, it’s obvious that ML is not really “intelligence” by any…eh…intelligent definition of the term. Instead, it’s just a set of powerful analytical techniques. Yes, those techniques do have the potential to transform entire industries. But only under our control. And only if we apply them well.
And so…all there is left to do in next month’s blog is to get round to a few of the rumours, concerns and even myths surrounding machine learning. Is it a locked black box with algorithms I can’t understand and therefore shouldn’t trust? Is it the same as robotics, or artificial intelligence? Is ML (or deep learning or AI) going to take my job? Or is it going to take over the world and kill us all? Elon Musk seems to think so, and he’s a smart guy, right?
Much still to talk about then. See you next month. Unless the robots have another plan.
1 I claim no ownership of this definition, obviously. I’m sure it’s been used a million times before. Someone may even claim to have come up with it.
2 Donald Rumsfeld’s famous “unknown unknowns”, you might say.
About the author:
David Socha is Teradata’s Practice Partner for the Industrial Internet of Things (IoT). He began his career as a hands-on electrical distribution engineer, keeping the lights on in Central Scotland, before becoming a part of ScottishPower’s electricity retail deregulation programme in the late 1990s. After a period in IT Management and Consulting roles, David joined Teradata to found their International Utilities practice, later also taking on responsibilities in Smart Cities and the wider Industrial IoT sector.