According to the Danish Maintenance Association (DDV), machinery maintenance is estimated to cost Danish manufacturing companies DKK 25 billions each year. Additionally, it is estimated that there is a globally cost reduction of $240 - $630 billions in 2025, thanks to the implementation of predictive maintenance [1].
This indicates that the potential is huge within reducing unplanned downtime and the consequential costs. This potential spans several industries ranging from energy to manufacturing where securing a high uptime of critical components is important. Most importantly it does not need to take years to reap these benefits but can be done within months.
Predictive maintenance describes a specific maintenance approach of which you predict future breakdowns, or the likelihood of a breakdown occurring, based on patterns in the data collected from the machines. This could for instance be pressure, flow, power usage, temperature, sound, and vibration of the equipment. These types of operational data can be combined with maintenance data to predict if a service interval can be postponed or be moved up.
The Possibilities
With a good prediction model, estimating the remaining lifetime of an equipment, makes it possible for production to plan the purchasing of spare-parts as well as planning the maintenance, such that it will not affect the planned production hours. This is a game changer from doing reactive maintenance where the maintenance crew never know what issue will arise next. Using predictive maintenance will reduce total costs not only on unplanned downtime, but also reduce the costs of spare-parts. It reduces the costs of spare-parts and maintenance hours as it runs with variable intervals following the mantra “if nothing is wrong with the bearing why change it?” For some companies, being able to detect sudden failures minutes before the breakdown occurs, can prevent dangerous situations leading to a safer production environment as well as reduce the probability for longer downtime periods. For others it can save them large amounts of money as batches in production will have to be thrown out if something goes wrong with a machine.
The Challenges
It requires a certain level of maintenance maturity before the values of predictive maintenance can be unlocked. Additionally, it requires the right data. Right data is when operational and maintenance data is sampled in the right sample rate, from the right sources, is representative, and is trustworthy.
Roughly put, predictive maintenance can predict two kinds of breakdowns: 1) sudden failures, and 2) breakdowns occurring due to wear-and-tear of an equipment.
If a company wishes to detect the sudden failures, an anomaly detection model can be implemented within months. Anomaly detection (or outlier detection) detects in shorter time spans, minutes to hours, before a breakdown occurs, and has a hard time spotting breakdowns occurring due to wear-and-tear. In order to be able to detect sudden breakdowns, the sample rate of data additionally has to be high. Once an hour or once a day is not sufficient when relying on predictive maintenance.
However, if a company wishes to estimate the remaining lifetime of an equipment, a larger dataset is required, of which you have experiences several (5-10<) breakdowns occurring due to wear-and-tear. In the blog post about predicting breakdowns on a water pump dataset, we experienced 7 breakdowns within approximately 6 months. Thus, the dataset was sufficient to create a machine learning model for both detecting sudden failures, as well as the remaining useful life of the water pump. However, in other cases the lead time before experiencing sufficient breakdowns to predict the wear-and-tear can be long.
Conclusion
The potential of using predictive maintenance to obtain a high operational up-time, and reduce total operational costs, is high. Independent on whether your company wishes to detect sudden failures or the remaining life of an equipment, you must retrieve the right data.
// Maria Hvid, Machine Learning Engineer @ neurospace
References
[1] Egedorf, S. (n.d.) Prædiktiv Vedligehold lover positivt i forhold til at kunne spare virksomheder for unødvendige udgifter ved at optimere vedligeholdelsesstrategien, og tilgangen anvendes i flere virksomheder. (DDV).
