Flexible Degradation Modeling via the Integration of Local Models and Importance Sampling

Sensors are generally used to monitor the degradation status of various engineering systems and units. Previous studies have developed parametric models to describe the degradation status of a unit, predict the remaining useful life (RUL), and make maintenance decisions based on collected sensor signals. However, the models above are parametric and are often restricted by assuming a specific functional form of the sensor signals. This study proposes a flexible model with fewer assumptions, which can be potentially adopted in various practical situations with static operating conditions. On the basis only of historical units with the most similar degradation trends, we aim to predict accurately the RUL of an in-service unit. First, we extract features that characterize the degraded progression of each sensor signal and define the distance between units based on these features. Then, we construct local models to describe the relationship between the extracted features and the failure time for RUL prediction. However, limited sensor measurements from the in-service unit may cause a high variance of features and lead to unreliable predictions. To address this issue, we integrate the importance sampling (IS) technique to incorporate population-level information, thereby improving prognostic performance. Furthermore, we use degraded aircraft engines for a case study to validate the proposed model using two datasets generated by C-MAPSS.