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

This article develops a supervised tensor dimension reduction-based prognostic model designed to predict an asset’s time to failure (TTF) using incomplete degradation imaging data. Unlike existing models that rely on unsupervised dimension reduction and require complete image streams, the proposed method integrates historical TTFs to supervise the extraction of low-dimensional tensor features from high-dimensional, possibly incomplete, degradation image streams. The model constructs a prognostic framework based on (log)-location-scale (LLS) regression, with parameters estimated via a block updating algorithm that guarantees convergence and admits closed-form solutions when TTFs follow normal or lognormal distributions. Validation through simulated heat transfer data and real-world rotating machinery infrared image streams demonstrates that the supervised approach consistently outperforms unsupervised tensor dimension reduction benchmarks across varying levels of data missingness, highlighting the benefit of incorporating failure time information in feature extraction for prognostic accuracy.