Remaining useful life prediction using a stochastic filtering model with multi-sensor information fusion

Abstract

As an important element in condition-based maintenance (CBM) of complex engineering systems, remaining useful life (RUL) prediction has attracted much attention over the last decade. Since the degradation process of a system is uncertain over time as being well recognized, the resultant RUL is random in nature and the uncertainty needs to be quantified in the prediction. It is well known that higher uncertainty means lower accuracy in life estimation and higher costs in subsequent maintenance actions. Hence one practical problem is how to reduce the estimation uncertainty of RUL for more accurate prediction. Another known fact is that multiple sensor information co-exists in practice and how to best use this multisourced information to achieve a better RUL estimation is a question not well answered. This paper proposes a stochastic filtering model for RUL estimation with available multi-sensor information. Firstly, a stochastic filtering model based on the Bayesian theory is used to describe the RUL distribution for each single sensor. This RUL distribution represents the conditional probability density function (PDF) of the RUL conditional upon the monitoring information history of the sensor. Then centralized and distributed information fusion approaches are applied to combine the multi-sensor information to form a new RUL distribution with the objective to minimize the RUL prediction uncertainty. Centralized fusion focuses on combining the multi-dimension measurements before the RUL estimation and then to predict the RUL based on the fused measurement. Distributed fusion, on the other hand, fuses the predicted RUL distributions of all sensors after obtained the RUL prediction based on each sensor information. Finally, the stochastic filtering model with multi-senor fusion information is applied for a case study to demonstrate the performance of our approaches with the prediction based on a single sensor. The result shows that both fusion methods outperform the model with a single sensor in terms of prediction accuracy.