Predictive algorithms make up a huge number of IoT success stories today. Generally applied as a way to predict failures or as a warning to provide maintenance in the near future, companies ranging from elevator makers to oil refineries are using predictive algorithms in their operations. But it’s come to my attention that not all predictive algorithms are the same, and figuring out what each of them are comprised of offers a window into how IoT is maturing in the industrial setting.
While IoT has become all the rage in the last decade, factories of every stripe have had continuous monitoring in place to detect equipment failures for much longer. In such situations, companies would apply physics to the problem of predictive maintenance. (If a piece of metal has x tensile strength, then if temperatures exceed y it will bend or break.) But now companies have the option of adding much more data and applying machine learning to an algorithm to create what are known as data-driven algorithms.
In today’s hyped-up world of PR pitches and buzzwords, data-driven algorithms are what people are trying to talk about when they talk about predictive maintenance. But not every situation requires data-driven models. And if you are going to use a data-driven predictive algorithm, it’s worth understanding some of the drawbacks. The use of data-driven algorithms vs. physics-based algorithms was a big topic at the Emerson event I attended in September, and was the focus of a conversation I had recently with Philipp Wallner, an industry manager with MathWorks.
At this week’s event, Peter Zornio, the CTO of Emerson’s Automation Solutions, spent a significant chunk of time explaining the difference between the two model types. He said that one of the bigger challenges for customers was figuring out when to rely on rules-based algorithms as opposed to machine learning. For example, predicting failures using machine learning can be tough when there isn’t a lot of existing failure data with which to train a model. In that case, relying on rules and physics makes more sense.
Ironically, when it comes to using machine learning vs. traditional physics-based algorithms, customers found developing AI-based-based algorithms a mere 2% more challenging than developing physics-based algorithms, according to a Software AG survey from earlier this year. Sean Riley, global director of manufacturing and transportation at Software AG, told me the results surprised him, mostly because more than 60% of the survey respondents said they found defining physics-based algorithms — or what the survey referred to as “threshold-based rules” — to be as difficult as integrating IT systems and IoT sensors into existing control systems. In other words, building rules with AI or using established parameters is hard no matter how you do it.
So I spoke with MathWorks’ Wallner to get his perspective. MathWorks makes two software products, MATLAB and Simulink, that have been turning data into statistical models for customers for more than three decades. Wallner made the case for both a physics-based and a machine learning-based approach, pointing out that in many industries the combination of physics-based data and an understanding of likely failures means that data scientists can simulate failure data to train an AI-based model.
He also pointed out that an understanding of physics-based behavior can lead to the creation of so-called virtual sensors, which can provide a more accurate understanding of what’s going on in the factory without needing to add a new piece of hardware. For example, measuring the electrical current used by a centrifuge could help create a virtual sensor that can determine the viscosity of the fluid inside and the subsequent wear and tear on the components driving the machine. Such virtual sensors are a combination of both physics and data-driven models.
This may seem like a persnickety difference, but given all the hype of AI and predictive maintenance, it’s worth looking at solutions that have already been tested and require less computing power as a first step. Training a model on a cluster of energy-guzzling GPUs could be akin to using a sledgehammer to hang a picture — ultimately unnecessary, and with the potential to cause further harm.