Diaz-Gonzalez, Abel; Coursey, Austin; Quinones-Grueiro, Marcos; Kulkarni, Chetan S.; Biswas, Gautam. (2026).听.听Mechanical Systems and Signal Processing, 261, 115026.听
Predicting the remaining useful life (RUL) of industrial equipment is important for planning maintenance and preventing failures. Many data-driven methods predict RUL directly from sensor measurements without explicitly modeling how equipment degrades over time. This study proposes a framework that instead separates the process into two stages: estimating the system鈥檚 current health and predicting its future degradation. The method uses sensor data to estimate health, continuously updates uncertain degradation parameters as new data become available, and produces a range of possible RUL values that reflects this uncertainty. A neural network helps adapt the estimation process while preserving uncertainty information. Testing on the N-CMAPSS benchmark showed competitive RUL prediction accuracy and statistically grounded uncertainty estimates across operating conditions, supporting the potential benefits of separating health estimation from future prediction in data-driven prognostics.
Fig. 1.听Overview of the proposed estimation鈥損rediction framework. In the estimation stage, the system condition and degradation-related information are inferred from the available observations. In the prediction stage, performance-specific degradation models are used to obtain the predictive distributions of the times at which each health index reaches a performance level聽饾憼鈭圼0,1]. These distributions are combined with the system threshold function and the desired confidence level to compute RUL predictions and associated confidence intervals.