There are some very good reasons to be interested in predictive maintenance. Think of the time saved! The costs cut! The resources better allocated! The more efficient you become at identifying which part, machine, or process is likely to break down and when, the more reliably you can reap the benefits.
But how reliable is “reliable,” really? How certain are these enticing benefits? As the hype around predictive maintenance has grown and grown, so has the obscurity. The (very real) challenges to achieving predictive maintenance have been downplayed just as the thing itself has been hailed as a cure-all. But nothing so unambiguously good can be entirely true. With predictive maintenance, as with everything else in life, there’s no free lunch; like other elements of “smart manufacturing,” it has to be approached soberly, carefully, and with full awareness of what it solves and what it doesn’t.
In this blog, we’ll run through the five most important things about predictive maintenance that no one wants to admit. Equipped with a better understanding of just what kind of future we’re being promised, we’ll be better able to meet it with a working strategy.
1. If It Were Just About Data, It Would Be Everywhere
Pundits like to wave their hands about the data-driven future of manufacturing. Of course, we at Dataiku believe very strongly in that future — but we’ll never wave our hands around the challenges and the dark spots. Yes, it’s true, the future is bright: Industry 4.0, Data Governance, Everyday AI. But without you, it’s just data.
If it were really all about data, and if everything from predictive maintenance to cross-functional process optimization magically followed from more accessible and intelligent data operations, then they’d be ubiquitous across the industry. The explosion of cloud storage technologies and services over the past decade has made data more accessible than ever before. And the rapid rise and democratization of AI and machine learning-enabled data processes have catapulted us into a new age of data manipulability and usability. Everyone has the data, and everyone can make use of it in new ways. And yet, predictive maintenance is far from ubiquitous.
That’s because it takes people to change processes.
2. Predictive Maintenance Is Nothing New
Not only is predictive maintenance not everywhere; but in some of the places where it does exist, it has been around for a relatively long time. Turbine manufacture, jet engine factories, and petrochemical plants, to take a few examples, have been using predictive maintenance for years. The same is true for some nuclear power plants.
The explosion in data accessibility and usability has, certainly, led many more firms and corners of the industry to embrace predictive maintenance. But it’s important to dispel the myth that it’s new, because the promise of the “new” often leads us to ignore the lessons and insights gained from past experience (the “old”).
3. Predictive Maintenance Is Not a Strategy
Those who get dazzled by hype also tend to miss the forest for the trees. Predictive maintenance may be promising, and it may even be the future, but it’s not a strategy. It’s a tactic.
To see this we only need to consider what our actual goal is and work backwards. For any manufacturing company planning for the future, the ultimate goal is to reduce costs. Because equipment and everything to do with it (purchasing it, operating it, maintaining it, etc.) comprises a significant part of any manufacturer's costs, the best strategy is one that performs maintenance as needed, taking into account cost, criticality, and capability.
Predictive maintenance is a useful tool here, but it has to be evaluated as a possible tactic among many in your planned strategy. For one thing, a given manufacturing operation might not actually deal with many failures on a regular basis. In that case, predictive maintenance would be most useful not as a way of predicting failures, but of predicting the problems that cause an increase in unplanned maintenance.
Cost is important, too. Predictive maintenance relies heavily on sensors that regularly provide information on equipment condition. And there is a price to adding and maintaining such sensors — every additional sensor incurs a charge from the original equipment manufacturer. There are also labor costs to consider: someone has to ensure that the sensors are calibrated and functioning properly, and that someone, of course, has to be paid.
The cost of predictive maintenance is most justified, then, if your equipment fails randomly, is critical to your production goals, and/or has high a cost-of-failure.
4. You Can’t Predict What You Don’t Understand
When we say that without you, it’s just data, we mean that processes only produce value if the person designing and benefiting from them understands what’s being processed. When it comes to predictive maintenance in particular, there will be nothing to predict if the people running the process don’t understand failure modes or operational theory. The point isn’t to shame anyone about a lack of knowledge (after all, each of us lacks plenty of knowledge!), but just the opposite: to begin with a realistic accounting of what you (and your team) know and what you don’t, so that you can figure out what to learn.
When it comes to predictive maintenance, there are a few crucial variables that need to be defined and understood before the process can produce valuable information for your analysts. The first step is to understand the data you currently have as well as the functions of your sensors. What is being measured, in other words, and by what sensors, and how frequently, and in what manner — all of these are crucial questions. You can’t predict when a car will need an oil change by looking at the gas gauge; the wrong sensor, measuring the wrong thing, at the wrong intervals will give you the wrong answer.
The second step is to figure out what data is quality and what data is not. Separating signal from noise is the only way to establish a reliable channel of communication between your sensors and your models, and between your models and your analysts.
The third and final step is to ensure that exploration is possible. You’ll want your model to test scenarios and possibilities and return reliable results as to their respective viability. But this is only possible if you can dependably determine whether the data you have is sufficient to indicate patterns that predict failures when they inevitably crop up. Sensors are costly, and you likely won’t have every sensor it’s possible to install. So ask yourself: is what I have sufficient? When, where, and why do failures occur? You can only explore if you can routinely answer these questions.
Only after you’ve developed these three steps can you build and run your predictive maintenance model and expect results.
5. If You Don’t Generalize, You Won’t Scale
It’s a major accomplishment to set up predictive maintenance for a handful of new, relatively similar machines at a small factory. But what if that factory is part of a larger network of manufacturing spanning tens (if not hundreds) of sites, each of which uses different machines of variable age? If your model can’t adapt accordingly, your local success is likely to remain an anecdote.
A well-oiled predictive maintenance system will be able to generalize to larger and more varied inputs, from hundreds of motors and drives in a manufacturing plant to thousands of them. Add to this mixers, computerized numerical controls, ovens, robotics, heat exchangers and pumps, and much more. And for any given type of equipment, your model will need to account for different ages of machines, different models, and different manufacturers — not to mention different control systems. The type and amount of sensors will differ and so will the failure modes and fault detection mechanisms.
Your model should be able to handle 50 sensors just as well as 5. In other words, it needs to be highly flexible and adaptable to different influencing factors and control variables.
Is It Worth It?
In many cases, yes. But it’s important to know your strengths and weaknesses before committing to change.
Unplanned maintenance is costly. In some cases, it’s catastrophic and even deadly. By some estimates, unplanned downtime can cost plants upwards of $100 million (depending on size and sector). The desire to chip away at downtime and recapture some of that lost capital is understandable, even crucial. And predictive maintenance, as has been proven time and again this past decade, can certainly help with that.
It’s just not as easy to implement or operate as some make it seem, nor is it the only way you can impact downtime. Predictive maintenance, as we’ve discussed here, is an ideal that needs to be brought down to Earth: don’t confuse tactics with strategy, and don’t let hype obscure reality. Most companies can generate a lot of value with the data they are collecting right now. They should start there and put a process in place that lets reliability engineers create analytics that shed light on existing problems, illuminate failure modes, and pave the way for prediction.
Step by step, with a measured outlook on the future, predictive maintenance will come to seem more like the tool it is, and less like the panacea it’s too often promised to be.