PhD researcher Collin Drent from Eindhoven University of Technology has developed a learning model for the maintenance of expensive and complex machines, such as chip machines, CT scanners and wind turbines.
It costs 72,000 euros if a chip machine fails for an hour due to a defect. It is logical that companies will prevent such mistakes at all costs. They do this through preventive maintenance. This means that you detect possible errors in time so that you can intervene before it is too late.
Predicting the optimal time for maintenance (not too late, but not too early either) is not easy, if only because each device is different. PhD researcher Collin Drent found the answer in smart mathematical models and data, a lot of data. He recently received his doctorate cum laude from the Faculty of Mathematics and Computer Science.
Chip machines are not the only high-tech devices that can cost a lot of money (and sometimes even lives) to break down. Think, for example, of trains or planes, CT scanners in hospitals or offshore wind turbines. It is estimated that unplanned machine downtime costs companies worldwide around € 50 billion each year. About half of this is due to defects. That is a considerable amount.
‘It is therefore very important that you detect possible faults in critical components in good time so that you can repair or replace them before it is too late,’ says Collin Drent, a researcher at the Stochastic Operations Research group. ‘But companies will not intervene too early either: Parts are expensive and you want to use them for as long as possible.’
An expensive scanner
To determine the best time to intervene, Drent conducted research into the so-called IXR scanner from Philips. These extremely expensive CT scanners (see image) allow physicians to perform image-guided surgeries that are minimally invasive to the patient.
Drent quickly discovered during his research that the standard models of preventive maintenance do not work well in this case. “They assume that units are all the same, and that you can therefore predict with great certainty when certain parts will be replaced. But that is often not true. Every CT scanner is different and so is the way and where it is used. Think about the temperature or the humidity. ‘
Fortunately, these modern devices produce an enormous amount of data due to the many sensors with which they are equipped. ‘You can use this data to make your models even smarter. This way, you can make a specific prediction for each unit and part, even if you do not know in advance exactly what factors are influencing the aging process. ‘
Drent used two different methods for his analysis: Bayesian Learning and the Markov decision model. ‘It has two advantages: by combining the learning capacity of Bayesian Learning with Markov’s decision-making model, I was able to make my predictions even more precise. Moreover, these algorithms are very transparent. So you know exactly what is happening, why. In this respect, the traditional AI methods such as deep learningwhere the exact function of the algorithms remains hidden in one black box. ‘
In the end, the researcher was able to reduce the maintenance costs of the IXR devices by an average of about 10 to 20 percent compared to the standard models. ‘And it’s really a lot when you consider how much money is involved in the maintenance of this type of high-tech equipment: the maintenance costs of such machines are generally at least as high as the purchase costs.’
In addition to CT scanners, Drent has looked at five other scenarios. It showed that his models in principle also work for e.g. wind turbines and chip machines. Drent calls his model one ‘united framework’.