Real-Time Predictive Analytics Using Degradation Image Data

Prognostics by using degradation image streams is challenging due to the high-dimensionality and complex spatial-temporal structure of images. In this paper, we proposed a penalized (log)-location-scale regression model that can utilize high dimensional tensors to predict the residual useful lifetime of systems. Our method first reduces the dimensionality of tensor covariates by projecting them onto a low tensor space. Next, the coefficient tensor is decomposed by utilizing CANDECOMP/PARAFAC (CP) decomposition. The decomposition facilitates further dimension reduction of coefficient tensors. By doing so, the number of parameters to be estimated is dramatically reduced. Furthermore, a numerical algorithm with global convergence property was developed for the model estimation.