Abstract

The goal of this project is predicting failures in a fleet of sterilizers deployed in hospitals all over the world. The characteristics of this problem are general to the field of predictive maintenance for different application fields. Companies are interested in predictive maintenance to reduce the down time of their machines. In general the list of critical components, whose unexpected breakdowns would result in stopping the machine, is long. Therefore, the scope of a predictive maintenance system should be predicting failures in a big number of different components. For several years, systems such as cars, sterilizers or industrial equipment have been equipped with a significant amount of sensors. Which signals to record is in general not decided based on the predictive maintenance needs, but on the requirements of security or controllers among other reasons. The sensors mounted usually don’t describe the particular behavior of the components of interest, but measure physical quantities that can be influenced by the different behavior of several components. Predicting what component will fail when requires historic data about the operation of the machines, but also needs to be linked to the occurrence of failures, so that we can label the recorded data. In general, companies have access and store data coming from their machines, but don’t necessary have access to the whole history of repairs. The owner of the machines can decide whether to perform maintenance and repairs with the official service or any other unofficial service. The main research goal of this project is to build a framework that allows predicting all type of failures that can happen in a machine.