Bayesian Online Changepoint Detection Algorithm for Degradation Monitoring and Prognostics of Electrical Devices

Abstract

Electrical devices that undergo frequent switching operations can suffer mechanical wear over numerous cycles. Identifying early signs of degradation through routine measurements is important for condition monitoring and predictive maintenance. This work applies Bayesian Online Changepoint Detection (BOCD) to multivariate signal data from an electrical device to detect the onset of degradation and estimate remaining useful life (RUL) under uncertainty.

Classical approaches such as rolling-window trend fitting react slowly to regime changes and require careful tuning. BOCD instead maintains an online posterior over the run length and recursively updates a predictive model as new observations arrive. We extend the classical formulation to a multivariate regression setting with operation-dependent trends, conjugate matrix-normal inverse-Wishart priors with the regression-coefficient precision calibrated from prescribed alarm bounds to confine initial predictions to the operational envelope, and a Weibull hazard function that encodes wear-out behavior. The resulting algorithm performs online inference without storing the full measurement history while providing calibrated predictive uncertainty.

The framework is first evaluated on synthetic degradation scenarios covering abrupt, gradual, stepwise, and accelerating fault modes, and then applied to real mechanical endurance-test data from a run-to-failure experiment. On the synthetic scenarios, abrupt changepoints are detected without delay and gradual degradation is flagged once the cumulative drift exceeds the noise level, with the lower, median, and upper RUL bounds converging consistently at end of life. On real endurance-test data, the algorithm maintains a stable RUL estimate throughout the healthy device life despite noise-induced changepoints, and provides a timely end-of-life warning once sufficient post-changepoint evidence confirms the degradation trend.