The precision of Predictive Maintenance

Mr Adrian, you support industrial customers in the area of production and maintenance in monitoring machine systems – what exactly can you contribute with your team as a Mathematician?

We want to make it possible to calculate the condition of machines. It’s about being able to make data-based predictions about when wear occurs and any parts which may need to be replaced. This replaces the prognosis models from the pure physics of the system: if I only consider that a metal has a certain life expectancy, but do not take other factors such as change in use or weather into consideration, I obtain less reliable results. However, if I record the respective current condition via sensors, it is possible to obtain clearer prognoses. For their calculation, we use machine learning to understand the dynamics of the systems on their data.

What do you need in order to determine the optimal time for Predictive Maintenance?

Mainly experience. The most important question is: how much data to I need? Do I have to go to the limit first and wait until something breaks, before I train a machine learning process with the data? From an academic perspective, you could claim that would be perfect, however, in reality you can’t do it and you wouldn’t want to. Predictive Maintenance provides an opportunity to extend maintenance measures based on data and evidence, rather than shorten the periods based on risk awareness.

What exact time is predicted?

Where machines are concerned, we refer to a wear margin: each system and each component has a margin which is reduced through its use and service life. What we cannot predict is the precise time at which the component will fail – there are too many influential factors to do this. But, we can stipulate a sort of best before date, from which the required quality can no longer be guaranteed. So there are customers who say, they will try operating the system for longer – and if it breaks much later then they say that the prognoses weren’t correct. But that is exactly what we want to avoid: the aim of Predictive Maintenance is always the greatest possible safety, so that is doesn’t fail or break unexpectedly. At the same time, the quality should remain consistently high. Not forgetting: wear often swallows up energy, for example, due to increased friction.

What challenges do you currently face in your research?

Where small components are concerned, such as sliding or ball bearings, we have understood the physics and we have it more or less under control: their technical data sheets generally show damage frequencies. However, if individual parts are compiled into something new, such as is the case in a machine or a trailer, there are a lot of additional influential factors, for example, speed, external temperature and the condition of the spare parts which were installed. This interaction makes each system individual. We can compare it to ourselves as human beings: we are all born with similar ‘equipment’, however, our bodies are influenced by many factors throughout the course of our lives.

How can we still reliably make predictions about when a spare part will be needed?

Real time monitoring is the supreme discipline here. Through monitoring, we can directly recognise changes and incorporate them in the prognoses.

What development opportunities do you still see in Predictive Maintenance?

The topic is definitely shaped by artificial intelligence. The progress through this technology centres on the fact that we are in a position to record functional and also non-functional dependencies with a great deal of influential factors using neural networks, without having to understand them as people. The only challenge is that they are highly individual. One thing these artificial neural networks also really like to do is learn things parrot-fashion. They act in a similar way to children at school – if they haven’t understood something, then they just learn it by heart. But then the networks don’t necessarily learn the correct aspects by heart. As Mathematicians, we therefore need to try to train such AI models for recurring circumstances, without having to learn them parrot-fashion. This also needs to remain within the limits of the computing unit: it always relies on PCs which only have a limited number of processes, cores and main memories. If they always have to keep adding more parameters, then of course that won’t work.

What potential do you see for technology in transport and logistics?

If semi-trailers are leased, the service life is generally part of the leasing contract. If a vehicle breaks down, then accordingly, the customer does not pay. It therefore makes total sense from an economical perspective, among others, to monitor the semi-trailer. It is particularly important in refrigerated trailers to monitor the cooling cycle and hydraulics – and to deliver a warning if it becomes apparent that an error is very likely to occur. The customer enjoys the advantage of a different payment model. While the manufacturer is in a position to remunerate warranties and quality in a different way, for example through extended warranties.

What do you personally find exciting about the topic?

That we always achieve added value. There is basically no reason not to start using Predictive Maintenance. I have never received data from a machine from which I could not derive interesting points, which manufacturers or operators had not yet taken into consideration. It’s not research in an ivory tower, you don’t have to simulate artificial data, rather we can really look at things precisely and experience lots of positive surprises: every day we see what is possible to increase efficiency and reduce costs.