Uncertainty-Aware Prognosis via Deep Gaussian Process

The problem of Uncertainty Quantification (UQ) is paramount importance when Machine Learning (ML) and Deep (DL) models are deployed in the real world. In context safety-critical applications, such as prediction Remaining Useful-Life (RUL) infrastructure industrial assets, relevance effective UQ approaches even higher, given potentially catastrophic consequences or substantial costs associated with maintenance decisions that performed either too late early. Particularly safety critical application contexts, transparency reliability essential requirements any ML-based solution not providing meaningful uncertainty estimates would fully satisfy desiderata. However, most ML DL techniques used for RUL estimation often designed to perform UQ, thus limiting their applicability real-world scenarios. To address this limitation, paper we investigate performance a recently proposed class algorithms, Gaussian Process (DGPs), predicting remaining useful lifetime (RUL). DGPs provide predictions, yet retaining expressive power learning capabilities modern methods. able scale very large datasets, which was limitation some previous algorithms but has increasingly become an requirement many applications and, particular, tasks key importance. main contribution thorough evaluation comparison several variants applied prediction. DGP evaluated on N-CMAPSS (New Commercial Modular Aero-Propulsion System Simulation) dataset from NASA aircraft engines. results demonstrate accurate predictions along sensible estimates, more reliable solutions (safety-critical) real-life applications.