Unsupervised Feature Selection for Characterizing Airplane Subsystem Behavior

The optimal characterization of operating conditions in aerospace subsystems is pivotal for ensuring optimal performance, safety, and reliability of aircraft. For any subsystem many sensor readings are obtained when a flight is in operation, selecting the best sensors (henceforth referred to as features) and studying those over time are key to understanding the behavior of the subsystem. This paper explores an unsupervised way of selecting the best features of one such subsystem CSAS (electric cabin supply air system) which can then be used to study predictive maintenance, fault isolation etc. A combination of centroid based, density based, and distribution-based clustering has been applied in this study. The features are a mix of unevenly distributed and mixed density clusters, so using any one of the clustering techniques gave inconclusive results. The selected features from the above clustering method was then compared with the feature importance scores of each of the sensors where CSAS output temperature and pressure were treated as the target variable. This comparison showed the results from the proposed clustering method was effective in finding the best features.