Particle filter-based damage prognosis using online feature fusion and selection

Damage prognosis generally resorts to damage quantification functions and evolution models to quantify the current damage state and to predict the future states and the remaining useful life (RUL). The former typically consists of a function describing the relationship between the damage state and a statistical feature extracted from the measured signals, thus the prognostic performance will strongly depend on the selection of a proper feature. Given the best feature may vary for different specimens or even at each time instant for the same specimen during damage progression, such selection is a challenging task but has received little investigation so far. In this context, this paper proposes a particle filter-based damage prognosis framework, which involves an online feature fusion and selection scheme. A prognostic model is considered for each feature, with a multivariate process equation, formulated using both a damage degradation function and a bias parameter, and a measurement equation linking the damage state and that feature considering a data-driven model and the bias. One PF is used to estimate the damage state, its evolution parameters, and the bias for each model. Then, at each step, the feature with the smallest estimated bias is selected as the best feature providing the most likely state vectors and is used to select the most likely samples of the damage state and growth parameters for predicting the RUL and for calculating the prior at the next step. The proposed prognostic framework is demonstrated by an experimental study, where an aluminum lug structure subject to fatigue crack growth is monitored by a Lamb wave measurement system.