Abstract
Online estimation of stiffness parameters from vibration data is a common building block of temperature-compensated damage detection. This study compares Bayesian filters for this task on a three-degree-of-freedom structure with a Bouc–Wen hysteretic element under varying temperatures: sequential importance resampling particle filters (SIR PFs), an action-aided particle filter (A-SIR PF) and its reformulations, a damage-factor PF, and extended, unscented and ensemble Kalman filters. Tracking, detection and localisation are assessed over paired Monte Carlo simulations of step and progressive degradation, at equal false-alarm rates. With the same model, the Kalman-type filters give the lowest errors, and adaptive systematic resampling brings the PFs to within 2–6% of them. Prior knowledge helps when it enters the transition model: a damage-factor PF, which includes a temperature–stiffness law learned from healthy data and the constraint that damage cannot increase the stiffness, is the most accurate filter, with errors on the healthy springs about 20% lower than those of the Kalman-type filters. The least-action principle, used as the objective of trajectory estimation, leads to a fixed-lag smoother that reduces the errors of the best filter by 25–62%, with a 1 s delay. At large thermal excursions, the environmental normalisation is necessary, and an autoencoder with standardised inputs and a conditional regression of the stiffness on the temperature are equally effective.