An Empirical Study on Linking DBSCAN Degradation Patterns to Kaplan–Meier Failure Probability and Maintenance-Demand Peaks in NASA C-MAPSS Turbofan Engines

Abstract

The aging of commercial and military aircraft fleets has produced an increasingly diverse mix of component degradation behaviors that pure demand-history forecasting is limited in anticipating. This paper presents a retrospective empirical study that links unsupervised discovery of late-life degradation patterns, nonparametric survival estimation, and association rule mining to characterize fleet-level maintenance-demand peaks for turbofan engines in a simulated replay setting. Using all four subsets of the NASA C-MAPSS dataset (FD001--FD004; 708 training units and 21 sensor channels per unit), we cluster engine degradation trajectories with DBSCAN on principal-component features, fit a Kaplan--Meier estimator on each cluster to obtain a per-pattern failure probability curve, and apply Apriori and FP-Growth association rule mining over (cluster, cycle-bin) co-occurrences with maintenance-demand peaks observed in a simulated 20-aircraft fleet. Across the four subsets, DBSCAN identifies 3--6 reproducible degradation patterns with cluster alignment above 0.78 against the training-RUL tercile reference labels. Cluster-conditional median failure times differ by up to 84 cycles within a single subset, and the resulting demand-peak forecasts achieve a mean F1 score of 0.67 against an RUL-only alarm baseline at 0.60. The improvement is moderate but stable across subsets, concentrated in early-warning windows of two to four flight weeks ahead of each demand peak.