Predicting Breakdown Risk Based on Historical Maintenance Data for Air Force Ground Vehicles
This work addresses maintenance scheduling inefficiencies for Logistic Readiness Squadrons in the Air Force, but it is incremental as it applies existing algorithms to a new dataset without introducing novel methods.
The paper tackled the problem of unscheduled maintenance causing downtime and increased costs for Air Force ground vehicles by developing a predictive system using historical maintenance data, finding that a Logistic Regression algorithm produced the most accurate results for optimizing maintenance schedules.
Unscheduled maintenance has contributed to longer downtime for vehicles and increased costs for Logistic Readiness Squadrons (LRSs) in the Air Force. When vehicles are in need of repair outside of their scheduled time, depending on their priority level, the entire squadron's slated repair schedule is transformed negatively. The repercussions of unscheduled maintenance are specifically seen in the increase of man hours required to maintain vehicles that should have been working well: this can include more man hours spent on maintenance itself, waiting for parts to arrive, hours spent re-organizing the repair schedule, and more. The dominant trend in the current maintenance system at LRSs is that they do not have predictive maintenance infrastructure to counteract the influx of unscheduled repairs they experience currently, and as a result, their readiness and performance levels are lower than desired. We use data pulled from the Defense Property and Accountability System (DPAS), that the LRSs currently use to store their vehicle maintenance information. Using historical vehicle maintenance data we receive from DPAS, we apply three different algorithms independently to construct an accurate predictive system to optimize maintenance schedules at any given time. Through the application of Logistics Regression, Random Forest, and Gradient Boosted Trees algorithms, we found that a Logistic Regression algorithm, fitted to our data, produced the most accurate results. Our findings indicate that not only would continuing the use of Logistic Regression be prudent for our research purposes, but that there is opportunity to further tune and optimize our Logistic Regression model for higher accuracy.