High accuracy flight state identification of a self-sensing wing via machine learning approaches

Motivated by the supreme flight skills of birds, a new concept called “fly-by-feel” (FBF) has been proposed to develop the next generation of intelligent aircrafts. To achieve this goal, Stanford Structures and Composites Lab (SACL) has developed a smart wing which embeds a multifunctional sensor network on the surface layup of the wing [1]. By leveraging structural vibration signals recorded from multiple piezoelectric sensors in the sensor network under a series of wind tunnel tests, data-driven approaches are developed to identify the flight state of this smart wing, i.e. angle of attack (AoA) and airflow velocity. Different preprocessing techniques are used including extracting 38 features in both time and frequency domains and standardizing the raw signals. Various supervised learning algorithms were applied to effectively establish the mapping from the feature space to the practical state space. In addition, it is found that 1D Convolutional Neural Network (CNN) can directly learn features from standardized signals and achieve similar performance to other algorithms using manually designed features. Compared with previous study [2], we have successfully achieved 96.55% test identification accuracy with the airflow velocity resolution improved from originally 3 m/s to 0.5 m/s under the same AoA.