Towards Prescriptive Risk Assessment: Utilizing Data Science to Model and Predict Top Degraders in Aircraft

SUMMARY & CONCLUSIONSMaintaining the readiness of aircraft remains a top priority for the U.S. military. Readiness is a function of spare part availability, reliability, and maintenance practices. Early identification of components and systems across the fleet in terms of these practices is key to maintain readiness. Historical approaches to identify top degraders remain largely reactive with a ranking of components once significant amounts of downtime have been recorded. Proactive and ultimately prescriptive assessment of top degraders will provide valuable lead time to decision-makers to reduce the impact of top degraders and overall downtime. This paper will discuss a data science-based approach to predict top degraders utilizing a military aircraft work order dataset. The approach achieved a high level of accuracy while maintaining a low false-positive rate. The unstructured text of the discrepancy narrative served as the main input into a neural network and machine learning models. Using natural language processing (NLP), the text was transformed into numerical value in which semantic similarity among words was captured. This allowed for phrases and word combinations in the discrepancy narrative to function as a key input into models. Numerical values of text combined with other work order attributes were revealed to have predictability on top degraders. Such findings implicate the significance of evaluating unstructured text in aircrafts’ discrepancy narratives on effectively modeling top degraders.The research indicated that it is feasible to utilize a combination of NLP and machine learning to develop a prescriptive framework focused on top degrader identification in terms of military aircraft. The results indicated that legacy aircraft maintenance data do hold valuable information that can be effectively extracted using NLP.The research further supported subject matter expert opinion that the majority of useful information is contained within the unstructured text. The analysis indicated that models only considering the unstructured text may be sufficient and may exhibit less variability in terms of results. This is potentially a major finding that can impact approaches to maintenance planning such as Reliability Centered Maintenance.