Zen and the Art of Ship Maintenance

(not all that is digital is gold)

I decided to write a brief article following the deafening noise about AI and maintenance, especially in shipping, an industry squandering millions in ineffective maintenance.

Bottom line upfront: digitization and IoT are extremely meaningful when it comes to implementing effective maintenance practices. On the contrary Big Data, High Frequency measurements, Machine Learning, and AI are largely irrelevant and a product of digital snake oil vendors.

Let's start with the current status of maintenance in shipping, which is largely based on scheduled maintenance (often mislabeled preventive maintenance): every so many running hours components A, B and C undergo maintenance, after so much many running hours components D, E and F, undergo maintenance... all the way to overhaul, rinse and repeat.

This maintenance strategy is ineffective, as the need for maintenance is not directly linked to running hours (not only), most of components are maintained ahead of when actually needed (higher costs), some fail prematurely (cost to repair the damage and loss of production hence operation disruption, hence further costs), some fail right after being maintained (maintenance induced failures, a not uncommon occurrence for auxiliary engines, btw).

The solution is, as strange is it might sound, the old adagio: "if it ain't broken don't fix it!". This does not mean reactive or corrective maintenance (i.e. replace components after they break) but maintenance on condition (only for components worth the trouble, we would still replace light bulbs after they fail): the operators monitor the health of the components and when their behavior starts deteriorating, also on the basis of how quickly it deteriorates over time, maintenance is scheduled and then carried out.

This is nothing strange and, at least conceptually, easier than it sounds: in principle we all should check every once in a while the treads of the tires of our cars: as the tread wears out we see how much margin is still there before we need to change them. Same as we try not to run out of coffee before buying a new pack. The fact that we do not check daily, or so seldom that we do not even do it any longer, is not by chance: most phenomena build over a long time, and this is why we do not need a magnifying lens checking in real time, as we drive, the condition of the treads and showing it in a dedicated portion of the dashboard.

In addition, engineering was not born yesterday, the mechanisms that lead to failures are generally very well known, as well as the symptoms of component degradation, and even more so the ones of impending failures.

We started down path of effective maintenance in the 1940's, with Condition Based Maintenance, CBM: we check the conditions of the equipment, and we perform maintenance when it is most cost effective.

We moved then to Reliability Centered Maintenance, RCM, which expands on CBM with a more systemic approach, looking not only at individual pieces of equipment, but also at their function within the plant, and at the cost effectiveness of different maintenance strategies for different components (which is why we do not do CBM on light bulbs!).

Finally, we landed on Predictive Maintenance, PM, which is a refinement of CBM: in addition to checking the condition of the equipment we predict when it would fail (or no longer provide the level of performance, safety, efficiency requested), so that we can perform maintenance even more cost effectively.

Going back to tires, CMB could prescribe to change them as soon as there is just 4 mm tread left, PM would require us check the rata of wear from one measurement to the next (X mm of tread wear of Y km travelled), calculate the wear rate, predict the future usage of the car over the next period of time (will be you going on a 10 thousand km road trip vacation or rest at your summer house, only driving to the local store to get groceries, meat for the grill, and beers?), and then assess whether you should change tires, and when, or you could keep using them only reassess their status at the next check.

All the above can be done with pen and paper, just as the good ole D&D, most likely AD&D for the maintenance nerds among us.

So, where and how digital tools fit in the above? Well, some do, some don't.

Digitization definitely does! As in all fields, the ability to store, retrieve, plot, and analyze data in a digital format is a huge plus. Even the basic haphazard collection of digital worksheets trumps pen and paper 6-0, 6-0, 6-0.

The next step is related to "smart" sensors, that measure and pre analyze the raw data into a form that is easily stored in your system and which is rich in information (e.g. translating vibration measurements from the time domain to the frequency one). These sensors can be hand held, we all started from there, fixed / cabled ones, IoT. IoT is an extremely nice to have for hard to access / remote pieces of machinery, it is a huge enabler, but, conceptually, it is much less disruptive than what it might sound.

Then we come to Big Data / High Frequency measurements. They are largely useless and, well, "Mostly Harmless" (cost aside).

Data might become very large if a system is very large, but this is not Big Data. Big Data is basically recording very many signals, mining most that it can be mined from an installation, regardless. When it comes to maintenance (and performance monitoring) it is largely irrelevant, the relevant signals in respect to the health of a component are well known and are relatively few, nth order effects are, well, nth order, therefore irrelevant by definition. Moreover, nth order effects are inevitably swamped by the noise which, is present in all measurements (and don't let me started by people pretending using AI to remove such noise and being able to identify nth order effects...)

Therefore going after Big Data in maintenance related applications almost inevitably means having little clue about the functional relationships between actions and reactions in a given systems. This is a huge organization issue, impossible to be solved digitally...

High frequency measurements are also largely useless when it comes to maintenance (and, again, also for performance monitoring). The reasons are that most phenomena relevant component status and failures take a relatively long time to develop, and then, often, lead to almost sudden catastrophic failures (e.g. crack propagation). In the former time frame, one good measurement per week, yes, per week, if not even less, is generally more than sufficient, after all one needs to check a long term / slowly evolving trend. In the latter time frame, the events happen too fast to perform any kind of action, so, again, high frequency is useless.

One caveat: all the above is unrelated to monitoring systems, automation, alarms, and safety (e.g. smoke detection systems) which feed on continuous monitoring of key parameters but very seldom, if ever, lead to the continuous recording of what measured / monitored.

Finally we come to machine learning and AI. Again, largely useless but no longer "Mostly Harmless."

The fact is that we are concerned about systems, very complex at times, of mechanic, electric and hydraulic components. The functional relations between the variables are largely known. These system are well behaved, non chaotic, there is no butterfly effect. IF, and it is a big if, granted, they are correctly installed and run in, they will run as expected, and will develop issues, in time, slowly, giving us time to plan appropriately.

We need neither ML nor AI to discover hidden patterns in respect to maintenance and failures, or about some form of "cross talks" between neighboring components (as the cross talk would show as a modification of the signal for a parameter already measured, e.g. external vibrations transmitted by the structure to a component).

In some systems, given the sheer size and complexity, one might argue that the cross talks could become relevant, therefore hidden patterns could be present. Even supposing this could be the case, every installation would be a case on his own, and the amount of data and time (and failures!) needed to train ML and AI would be such for this proposition to be unsustainable. In addition, also in this case, high frequency data would be useless as measuring the same variable every second does not create 360 times more information than measuring it every hour. Yes, the cost of doing it, regardless, could reduce to a manageable level, still the significance will be quite low. And good luck making sense of TB of largely useless data that might have to be examined / disclosed in court after an accident.

The tire wear and the prediction of the time to renewal needed neither ML nor AI, it was simple (at times not so simple) data analysis, which is definitely best performed by internally by a computer based systems, still, don't be fooled, it is neither needs ML nor AI.

But the real problem with ML and AI in maintenance is another, and it has to do with the human factor. It is the pretense that we can have an black box digital system (obscure in case of ML, a blackbody incase of AI) that can be exploited replace skilled operators with less skilled ones, and fewer (does anybody see a pattern in shipping?). This is accompanied by the myth that human factor is the largest cause of accidents (it is actually organizational failures, using humans as cheap and expendable scapegoats, and issues with man-machine interface operating in either non-congruent or undocumented ways).

If the operators do not understand how the pieces of machinery they operate can fail, and why they need to keep the machinery in good running conditions (not just the how, not just the when, not just the checklist, but real understanding), they should never run that machinery. And if they still do, and the machinery fail, it is not their error, but an organizational one.

Final words.

Civil aviation has massively invested in and profited from CBM and RCM way before digitization became a thing. There is absolutely no reason why shipping could not do and would not profit from ding the same.

Invest in crew, invest in quality ships, invest in maintenance, invest in "good" digitization. Theses are long term and high return investments.

Stay clear from digital snake oil vendors, who unfortunately abound.