Generalizing application agnostic remaining useful life estimation using data-driven open source algorithms

Independent on the industry, todays technical systems produce rich amounts of data. Cars, manufacturing equipment, and fridges have one thing in common: Repairing equipment after it broke causes additional costs due to, e.g., unplanned downtime, affects the products' customer image negatively and can even cause safety critical consequences (e.g., braking systems). Therefore, a strong need for prognostics that allows for replacing a component prior to breakdown is crucial. This inherently requires to estimate the remaining useful lifetime (RUL) of critical objects. Plenty of research has been conducted, trying to predict the RUL. However, most existing approaches share the following caveats: First, they are built and optimized for a particular problem (an therefore tend to over fit). Second, they are often based on proprietary software or closed source, preventing other researchers to use the algorithms out of the box. In this article, we compare four different algorithms on a variety of datasets: A 'naive' classifier that serves as a baseline, an algorithm that scored first on the most cited RUL prediction datasets, and two novel approaches. The first being 'bucketed random forest' which is able to predict the RUL accurately while requiring low computational effort once trained. Second, a downward scalable similarity-based approach, yielding comparably results while being very lightweight, which positively influences the testing time and model size. All algorithms in this article are open source and evaluated for general validity on a variety of different datasets.