In the realm of maintenance analytics, people talk a lot about physics-based vs. machine learning-based models. The old guard tends to favor physics, while those (often younger) folks with a penchant for data science tend to be more excited about machine learning. Let’s first set out to define the difference between these two paradigms, and then also take some time to break down the distinction between these two worlds to show that maybe they’re not so different…
In this article, I’m going to include some awesome perspective from a friend and business associate, Magnus Akesson, the Chief Information Officer for GE Power Manufacturing. Magnus runs IT and digital innovation across 50+ GE Power manufacturing sites, and he previously oversaw a similar program within GE Power’s repair facilities. He is truly a world class expert on power equipment, maintenance, manufacturing, and more – I tremendously value his insights!
Before we get to what Magnus shared with me, let’s quickly illustrate these concepts with simple examples. A physics-based model uses the laws of physics to make predictions, for example: the blade of a turbine will snap if we exceed a specific torque threshold. This is a very specific calculation governed by physics equations. The pressure on the blade is X, which creates a force of Y, multiplied by the blade radius is Z – and if Z exceeds a known shear strength threshold, the blade will snap.
Machine learning models instead focus on pattern recognition (as opposed to the laws of physics). Let’s recast the above example with machine learning models… suppose we notice a subtle change in the vibrational signature of the turbine, followed by a change in power output, followed by the blade snapping. If machine learning software can study enough examples of this, we can build a predictable model that correlates the vibrational and power patterns with this failure mode. We don’t have a perfect equation that relates vibration, power output, and blade snapping (nothing close to it) – but we’ve recognized a pattern that enables us to predict this outcome. The difference is known specific mathematical equations vs. learned pattern recognition.
Knowing that Magnus was an expert on both turbines and machine learning, I emailed him the example above and asked him for feedback. He replied with the following, which completely changed the way I thought about this concept:
So I have a bit of a “fundamental disagreement” on this one. In the end, Industry 4.0 machine learning is always physics-based (especially the mostly adopted branch of supervised learning). Machine learning is nothing other than function approximation, AKA y = F(x), except the way you do it is not what you learned in calculus. Modern algorithms instead have a unique way of iterating themselves to a mathematical function that approximate a complex non-linear relationship. Now this enables people to predict better outcomes based on the input variables you give the models. But because one can now do so using hundreds or thousands (or more) input variables, we think of it as pattern recognition and not physics… but it is and always has been physics!
Why am I saying this? Say when you apply a basic Decision Tree Ensemble (such as a Random Forest) on a predictive maintenance case, you use training data about temperature, humidity, vibration, current, amperage, pressure, etc to predict when the equipment will fail (also known as anomaly detection on time series data). This is nothing other than taking physical parameters and using those parameters to predict a future event. So it is inherently 100% physics-based. Fair enough, when you apply for a credit card and the bank makes a prediction about the default risk of you as a potential customer, they too use the same models and it is not necessarily a physics-based scenario. But in the digital industrial world, it almost always is.
In other words, it is not about specific equations/laws vs. patterns. It is definitely not about patterns vs. laws of physics. Indeed it is completely the opposite: it is about using pattern recognition to (non-linearly or linearly) approximate physical conditions that indicate a pending failure. So the point is more about how you can represent this immensely complex physical world through ever more advanced machine learning models that allow us to “see” relationships and patterns that weren’t previously possible. And how that in turn leads to a new way of operating a business, new skillsets needed to manage a physical process through statistical KPIs, and a new way towards capturing, processing, and managing data on the edge and in the cloud.
Now that’s pretty awesome stuff! And it’s really not practical to make an argument against Magnus’s points (so I didn’t try, despite the fact that he “fundamentally disagreed” with me – he’s right). The idea here is that the new age machine learning-based models are simply deeper physics-based models that were previously unfindable before computer-based data science horsepower. They are manifestations of the same thing.
Perhaps it would behoove stakeholders in the market to stop splitting hairs over the seeming difference between these two paradigms. I have been engaged in sales dialogs where potential customers rejected vendors because they leveraged machine learning-based models instead of physics models; I’ve also seen it the other way around. Hopefully the rift between data science believers and non-believers will begin to close, and someday we can all agree that the old ways and the new ways can converge to predict the brightest future for happy, healthy equipment all around the world!