skip to Main Content

Demystifying Machine Learning: Part I – Definitions

No machine is flawless. Given enough time, imperfections will develop, faults will progress and components will fail. If you are not prepared for this inevitable breakage, it can mean a sudden cease in production, the need for rapid (and therefore costly) repairs, and a large increase in O&M costs because of the need of a large stock of spare parts, just in case. If you are prepared however, the parts are repaired and replaced as part of routine maintenance, there is no unexpected loss of production, and O&M costs are minimised with multiple repairs grouped together. This is the essence of predictive maintenance – predicting when a failure is going to occur and planning appropriately for it.

 

Predictive maintenance doesn’t need to be complex; noticing that a generator is making a strange new repeating “thunk” sound can be good enough to give you some warning. However, making such an observation is often difficult in the wind industry, when your failing component is 100 meters in the air (and often several kilometers out at sea!).

Leveraging sensor data can dramatically reduce O&M costs – see this report on the use of AI in the industrial sector for more information:

https://mck.co/2GI6Y66

When you cannot be near your component, one must rely on a multitude of sensors to act as your eyes and ears. Rather than being alert for potential failures, a monitoring engineer will look for patterns in an increasingly complex array of data, searching for signals that indicate a potential failure is upcoming. This large and complex suite of data is an ideal problem to apply AI/machine learning techniques to. Studies on the use of AI in predictive maintenance are showing [[1]] asset productivity increases of up to 20%, and overall maintenance costs reductions by up to 10% over human capabilities alone.

 

DIFFICULTIES WITH DEFINITIONS

 

But what is “Machine Learning”? Ask this question to most industries, and you will likely get a broad range of answers: Some people think it will revolutionise every aspect of our lives. Others believe it will lead to scenes reminiscent of a Terminator film. Others still dismiss it as nothing more than a passing trend. All these views are probably wrong, but all of them have an aspect of the truth in them too. In any case, everyone today seems to have an opinion on this new technology.

 

Ask anyone to actually define machine learning however, and you will get a vague answer. These vague definitions are often centred on the idea of computers learning from examples, rather than being explicitly programmed. This learning definition has always felt somewhat cyclical; it defines “Machine Learning” as “Machines Learning”. More than that, it is unsatisfying, as it does not help anyone intuitively understand how the machine is coming to its conclusions. The phrase `black-box device` is often employed here, as if machine learning was an impossible to understand tool, one that humans should never even attempt to understand as the task would be fruitless.

Machine Learning is often described as a black-box system, where data is fed into to the system and answers extracted, with no way of knowing how the system is coming to its conclusions. Original image from https://www.machinecurve.com/index.php/2017/09/30/the-differences-between-artificial-intelligence-machine-learning-more/

 

I disagree with those who hold such `black-box` views. Going further, I think an intuitive definition of machine learning is not only possible, but is a more natural way to understand the potential of machine learning. Such a definition should make the following immediately apparent:

 

  • When I should be using machine learning to solve a problem?
  • What kinds of data will provide the greatest enhancement to my solutions?
  • How is it coming to those solutions, regardless of any particular implementation?

 

Before coming to this definition, it is worth spending a few paragraphs discussing why machines and humans have generally been suited to different kinds of tasks. This should set up the context for our intuitive machine learning definition.

 

COMPUTERS AND HUMANS ARE GOOD AT DIFFERENT TASKS

 

Computers have traditionally been very good at working with simple mathematical objects; things like numbers, simple shapes, databases. We would never ask a human to add one million numbers together; we could never trust the result. But we happily pass such a task to a computer and trust the results instantly.

As such, any sector dominated by simple mathematical objects have been taken over by computers. They are faster than humans, more accurate than humans, more reliable than humans, and they don’t need to be fed or sleep.

Humans, on the other hand, are masters at working with abstract ideas and patterns. This is no coincidence; such pattern recognition has kept us alive (think about the importance to your ancestor’s survival of recognising the shape of a lion or the pattern on a poisonous mushroom). Without even thinking, everyday we identify shapes that correspond to other humans and establish whether they are friendly or hostile, we flawlessly identify words and their meanings despite them only being pressure waves in the air, and we can identify complex objects in a busy scene without even trying.

Hericium Erinaceus, also known as the `Lions-mane mushroom’. Healthline.com suggests that such mushrooms “can be enjoyed raw, cooked, dried or steeped as a tea”. Original image https://commons.wikimedia.org/wiki/File:Hericium_erinaceus.JPG

 

None of the examples above can be encoded as simple mathematical objects, and so computers have traditionally struggled with these kinds of pattern recognition tasks (how would one write down the equation of a mushroom or lion? An impossible task.)

 

AN ACCESSIBLE DEFINITION OF MACHINE LEARNING

 

However, over the last 10-15 years, rapid developments have taken place within the machine learning and hardware fields, enabling computers to perform exactly these kinds of abstract pattern recognition. If we accept (for now; I’ll go into more detail in later blogs) these historical developments as the primary driver of machine learning, then a simple definition of machine learning drops out, without any need to allude to mathematics, specific algorithms, or implementations:

 

 

MACHINE LEARNING

A suite of tools and algorithms that enable computers to perform pattern recognition.

 

 

Whilst such a definition might not be revolutionary, and indeed even seem obvious in hindsight, it is suddenly trivial to answer the questions I posed above:

 

  • When should be using machine learning to solve a problem? When we suspect there is a pattern to be discovered in the data.
  • What kinds of data will provide the greatest enhancement to my solutions? Data which contains information relevant to the pattern we are trying to detect.
  • How is it coming to those solutions, regardless of any particular implementation? It is looking for patterns, identifying those patterns, then automating the detection of those patterns.

 

As an example, in predictive maintenance humans have often looked for typical patterns in the data that indicate a component is about to fail. Well now a computer can learn that pattern and automate that process. Indeed, it may even find new insights that the humans missed before; patterns that were there but were too complex or subtle for the human to detect, enabling it to detect such faults earlier than any human could have.

 

There is clearly a pattern here, but a human would struggle to identify it. From a mathematical perspective however, this huge complex pattern (a subset of the Madelbrot set) can be described simply by the equation zn+1 = zn2 + c. Original image from https://en.wikipedia.org/wiki/File:Mandel_zoom_07_satellite.jpg

 

Over the coming weeks and months, we hope to explore the further ramifications of this intuitive definition, examine in detail the kind of problems we can now solve by using machine learning and its pattern recognition, as well as how the wind industry is changing and developing alongside this new tool.

 

Stay tuned for Part II of this blog – We will explore specifically some of the further ramifications of this definition of machine learning and look at how it immediately guides us to best practices when implementing our own machine learning solutions in the wind industry.

Back To Top