Data-Driven Flight State Identification via Time-Series-Informed Features and Convolutional Neural Network

In this study, a data-driven approach to identify the flight state of a bio-inspired self-sensing wing is proposed. The flight state is characterized by the angle of attack of the wing, and must be inferred from vibrational signals recorded on the surface of the wing. Various transformations of the original vibrational time-series are identified and used as features for the novel multi-branch one-dimensional convolutional neural network that is specially adapted to multiple input time series which may have different dynamics that indicate flight state. We find that the additional time-series transformations improve classifier performance. We also find that the classifier is suitable for online-classification, and is very capable of stall state identification, thus providing a new high-accuracy method for flight-state identification in the next generation of intelligent self-sensing aircraft.