Empirical Evaluation of the Impact of Data Pre-Processing on the Performance of Predictive SHM of Jet Engines
Abstract
We evaluate the impact of data pre-processing on the performance of predictive Structural Health Monitoring algorithm on a real case study involving dozens of jet engines. A simple robust four-step framework is designed to this effect, made of 1) outliers removal, 2) range scaling, 3) variable selection (either by Òmanuallyî evaluating variable correlations or by quantification of variable importance via random forests) and 4) evaluation of the predictive performance of a unique selected binary classifier (random forests). The results contrast with the intuition and the literature, since pre-processing raw data decreases predictive performance in half of the cases analyzed. The isolated influence of each of the pre-processing techniques rank in this order: important variables chosen through random forests has the highest positive impact, followed closely by variable scaling and outlier removal to a lower extent, while the Òmanualî variable selection via the correlation matrix exerts a slightly negative impact on predictive performance. The influence of combining pre-processing techniques is in line with the isolated influence of each technique. However, a detailed evaluation should be done for every application since these results might be due to the high data quality of aerospace engines or to the characteristics of random forests.
Origin : Files produced by the author(s)
Loading...