Resumen de Simplified models of remaining useful life based onstochastic orderings

Luciano Sánchez Ramos , Nahuel Costa Cortez, Inés Couso Blanco

• We introduce a method for designing simplified models of Remaining Useful Life (RUL) that is especially suited for deployment in resource-constrained environ- ments. Instead of accurately predicting the RUL via complex nonlinear functions, our approach jointly learns (i) a probabilistic health state model and (ii) a stochas- tic ordering function — represented by a monotonic neural network — so that the resulting health indicator is comonotonic with the true RUL. Notably, this work is the first study where the learning task simultaneously optimizes both the predictive model and the criterion for comparing model quality. By co-optimizing these ele- ments, our method selects the simplest representation that preserves the essential ordering of degradation, as measured by a smooth approximation of Kendall’s τ statistic. Experiments on the CMAPSS benchmark and real-world datasets (in- cluding turbofan engines and road tunnel fans) demonstrate that our approach achieves competitive prediction accuracy with a drastic reduction in the number of parameters.