An adaptive functional regression-based prognostic model for applications with missing data

• We model degradation signals with missing data with the goal of predicting remaining lifetime.
• We examine two types of signal characteristics, fragmented and sparse.
• We provide framework that updates remaining life predictions by incorporating real-time signal observations.
• For the missing data, we show that the proposed model outperforms other benchmark models.
• For the complete data, we show that the proposed model performs at least as good as a benchmark model.

Most prognostic degradation models rely on a relatively accurate and comprehensive database of historical degradation signals. Typically, these signals are used to identify suitable degradation trends that are useful for predicting lifetime. In many real-world applications, these degradation signals are usually incomplete, i.e., contain missing observations. Often the amount of missing data compromises the ability to identify a suitable parametric degradation model. This paper addresses this problem by developing a semi-parametric approach that can be used to predict the remaining lifetime of partially degraded systems. First, key signal features are identified by applying Functional Principal Components Analysis (FPCA) to the available historical data. Next, an adaptive functional regression model is used to model the extracted signal features and the corresponding times-to-failure. The model is then used to predict remaining lifetimes and to update these predictions using real-time signals observed from fielded components. Results show that the proposed approach is relatively robust to significant levels of missing data. The performance of the model is evaluated and shown to provide significantly accurate predictions of residual lifetime using two case studies.