A Supervised Tensor Dimension Reduction-Based Prognostics Model for Applications with Incomplete Imaging Data

Abstract

Imaging data-based prognostic models focus on using an asset's degradation images to predict its time-to-failure (TTF). Most image-based prognostic models have two common limitations. First, they require degradation images to be complete (i.e., images are observed continuously and regularly over time). Second, they usually employ an unsupervised dimension reduction method to extract low-dimensional features, and then use the features for TTF prediction. Since unsupervised dimension reduction is conducted on the degradation images without the involvement of TTFs, there is no guarantee that the extracted features are effective for failure time prediction. To address these challenges, this article develops a supervised tensor dimension reduction-based prognostic model. The model first proposes a supervised dimension reduction method for tensor data. It uses historical TTFs to guide the detection of a tensor subspace to extract low-dimensional features from high-dimensional incomplete degradation imaging data. Next, the extracted features are used to construct a prognostic model based on (log)-location-scale regression. An optimization algorithm is proposed for parameter estimation, and closed-form solutions are derived under certain distributions. Simulated data and a real-world data set are used to validate the performance of the proposed model.