Real-Time Bayesian Modeling for Industrial Condition-Based Maintenance Applications

Abstract

Condition-based Maintenance (CBM) has proven effective in minimizing unnecessary downtime by scheduling maintenance actions only when indicators suggest machine health degradation. However, much of the existing CBM literature focuses on classification algorithms using benchmark or static datasets, leaving a gap in real-time modeling with realworld data model evaluation. This study addresses this gap by proposing a Bayesian Regression Model combined with the Bayesian Online Change-Point Detection (BOCD) algorithm, applied to regression residuals, as a heuristic for modeling machine states and transitions based on real-time vibration and temperature signals. The approach is based on two hypotheses: (i) the Bayesian regression model captures the current machine state, and (ii) residual-based change-point detection identifies critical state transitions with enough time for a machine maintenance or replacement action. Validation is performed using a proprietary dataset from a Brazilian startup, collected through battery-powered sensors monitoring vibration and temperature from rotary machines. Results confirm the effectiveness of the proposed method, demonstrating its promising application for the CBM scenario as an accurate real-time machine condition monitoring.