Predictive or condition-based maintenance (CBM) promises us the moon. By monitoring certain conditions (such as vibration, temperature and the quality of lubricants) we presumably can predict forthcoming equipment problems and prevent these from happening. Predictive maintenance would also lower the overall maintenance cost as it reduces the need for regular, preventative maintenance. Consequently, it would also increase the availability of the equipment. What more could we ask for?

Expensive, I hear you say? Impossible? If we may believe all gurus, the IoT and Big Data will be the answer to all of it. And at relatively low cost too. Well, get to it, then.

It was announced only recently, for instance, that the well-known Erasmus bridge in Rotterdam will be equipped with sensors as well. You can read the whole story here, in Dutch: http://www.ad.nl/rotterdam/sensor-in-strijd-tegen-brugstoring~a1cdee1e/. These sensors will monitor the state of an electric motor and the component it actuates. It should allow maintenance to act before it is too late and thus prevent unnecessary traffic issues in the city due to problems related to the bridge.

Columns and blog posts like these also exist to spark the discussion and voice some criticism. So, hear me out, and let me know your thoughts.

So, back to this predictive maintenance thing, would it really be that easy? On further reflection, I do see some obstacles on the road. Because predictive maintenance is based upon a predictive model; a model that tells us “if this, then that…”. And this has quite some implications.

Malfunctions Desired?

First, it implies that I have historical data about both the factors (conditions) that can lead to problems as well as on the actual occurrence of those problems: without problems, no model. I cannot relate conditions to results. And if I only have a very limited number of occurrences, the model will be very unreliable to say the least. It is interesting to read that in the article, the supplier of the system (Semiotic Labs) comes to the same conclusion: “When there aren’t any malfunctions, we can’t measure”. So, all very trendy all this IoT and Big Data stuff, but without any problems it won’t be of a lot of help I guess.

Next, we don’t start off with a clean sheet either of course. Companies typically already have some sort of (regular preventative and/or autonomous) maintenance program in place. This already prevents quite a lot of problems in the first place. So how to get to a reliable, predictive model in such case? Stop the current maintenance activities just to see what happens?

Correlation or Causality?

Or we could turn to historical data and use statistics. But as you all know very well, “results from the past are not a guarantee for the future”. It makes me think of the use of Technical Analysis (TA) in improving your stock trading results. Tell me, who has made millions using TA? Or does the future always hold something else in store for us?

Maybe try something else: causality instead of correlation. But this implies the use of subject matter experts that are able to explain how one thing leads to another. And that they can formalize these relationships so they can be built into a causal model. My guess is that many of these experts will say something along the lines of: “it all depends…”. And that multiple conditions and interactions between these conditions may lead to the potential occurrence of the problem under investigation. But in any case, it won’t be an easy task to formalize such a model. Most models I have come across therefore, are stochastic models.

And one should understand that not only equipment conditions matter. Environmental factors also come into play. There is a reason that sellers of appliances, for instance, often explicitly exclude certain cases in their warranty conditions. And if we can establish that there are such environmental factors, how are we going to monitor all of these then?

I vividly remember a study done at ASML, the largest supplier in the world of photolithography systems for the semiconductor industry, done by the Eindhoven University of Technology. The study revealed that the causal, expert-based model was the least reliable of all the models tested on the quality of their predictions.

Both the study and the article about the Erasmus bridge also came to the conclusion that monitoring one or only a few factors will not prevent a complex piece of equipment from breaking down. “We concentrate ourselves on the electric motor, but a bridge can break down in many more ways. We don’t have a solution for that”.

To sum it all up, not an easy task to get to a predictive model. Even when the IoT and Big Data will give us enough data, getting to a predictive model is a whole other story…

Dealing with Uncertainty

And even if in the end we get to a predictive model, we should never forget that it is a model only. It means that there is a certain probability its prediction might be wrong. The outcome of the model is surrounded by uncertainty. “It might possibly go like this, but we’re not certain”. We humans often react to these uncertainties in the same way, particularly if there are high failure cost involved, viz. “better safe than sorry”. And if the failure cost is negligible and the model complex, cursed with uncertainties and difficult to create, we should maybe ask ourselves the question if predictive maintenance is worth looking at anyway. Put differently, if you don’t mind, I think I’d rather take my umbrella just to be sure…

Golden Egg or Air Bubble?

So, the question is: what are we really going to gain from predictive maintenance? Are we really going to invest in predictive models and systems? Will these models ever get to the level of reliability needed to be more cost-effective than a simple, regular preventative maintenance strategy? Will predictive maintenance in the end turn out to be the golden egg or the air bubble? I’m honest, despite all the buzz, I don’t know yet. But maybe you know more about the future than I do?