Remaining useful life estimation in aeronautics: Combining data-driven and Kalman filtering

Data-driven prognostics can be described in two sequential steps: a training stage, in which, the data-driven model is constructed based on observations; and a prediction stage, in which, the model is used to compute the end of life and remaining useful life of systems. Often, these predictions are noisy and difficult to integrate. A technique well known for its integrative and robustness abilities is the Kalman filter. In this paper we study the applicability of the Kalman filter to filter the estimates of remaining useful life. Using field data from an aircraft bleed valve we conduct a number of real case experiments investigating the performance of the Kalman filter on five data-driven prognostics approaches: generalized linear models, neural networks, k-nearest neighbors, random forests and support vector machines. The results suggest that Kalman-based models are better in precision and convergence. It was also found that the Kalman filtering technique can improve the accuracy and the bias of the original regression models near the equipment end of life. Here, the approach with the best overall improvement was the nearest neighbors, which suggests that Kalman filters may work the best for instance-based methods.