Predictive maintenance with Bayesian deep learning

Nabergoj, David (author), Štrumbelj, Erik (mentor)

We assess the usability of Bayesian deep learning methods for remaining useful life estimation. To compare Bayesian models to standard deep learning models, we propose a model evaluation method based on simulated device maintenance. We find that Bayesian models outperform their architecturally equivalent deep learning models on synthetic data as well as on two benchmark datasets. The proposed evaluation method is relevant for practical applications and research, as it directly estimates maintenance costs and allows for more easily interpretable model comparisons.

• Language: English
• Keywords: predictive maintenance, remaining useful life, Bayesian deep learning, model evaluation
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2021
• PID: 20.500.12556/RUL-133124
• COBISS.SI-ID: 84824579
• Publication date in RUL: 12.11.2021

Secondary language

• Language: Slovenian
• Title: Napovedovalno vzdrževanje z Bayesovim globokim učenjem

Abstract

V tej nalogi ocenimo uporabnost metod Bayesovega globokega učenja za ocenjevanje preostale dobe koristnosti naprav. Predlagamo metodo za vrednotenje modelov, ki temelji na simuliranem vzdrževanju naprav. Na podlagi eksperimentov z umetnimi podatki in dvema referenčnima podatkovnima množicama ugotovimo, da so Bayesovi modeli boljši kot standardni globoki modeli z enako arhitekturo. Predlagana metoda za vrednotenje je relevantna v praktičnih aplikacijah in raziskovanju, saj neposredno oceni stroške vzdrževanja in omogoča interpretabilno primerjavo modelov.

• Keywords: Napovedovalno vzdrževanje, preostala doba koristnosti, Bayesovo globoko učenje, vrednotenje modelov