Scheduling of preventive maintenance using prognostic models - A case study on elevator doors

Aim of this thesis is to research how maintenance process of elevator doors can be optimised. To this end, a business goal is set. It is to decrease the amount of unplanned maintenance visits caused by door malfunctions. A corresponding analytics goal is defined, which is used for ranking elevators that would make best candidates for maintenance in this respect. Thesis shows that scheduling of preventive maintenance visits using prognostic models would be beneficial in reducing the rate of unplanned maintenance visits. The result was obtained via a practical case study using a real-life dataset containing data for more than 20 thousand elevators for a period of two years. Major part of the work was to form this dataset from (partially incomplete) raw data that consisted of various maintenance records and condition monitoring data on elevator doors. Tested models were the well-known Cox Proportional Hazards model, and a more recent recurrent neural network model called Weibull Time to Event RNN (WTTE-RNN). Survival analysis methods were used to extract information from partial observations. Prior to training the models, stratified survival curves were obtained via Kaplan-Meier estimator for two groups: all elevators and freshly maintained elevators. Difference in these curves quantifies the difference that preventive maintenance visit generally yields. Then, prognostic models were used for producing daily predictions of the survival curve for each elevator. Elevators were ranked using these predictions and based on how much their condition seemed to have worsened over time (as this is thought to capture their potential to benefit from maintenance). Finally, a comparison between highly ranked elevators over all elevators is provided to demonstrate the skill of the models