Modeling aircraft similarity with musical auditory feature extraction

Abstract

The increased use of unmanned aircraft systems for drone delivery and air taxis means the psychological effects of these systems (annoyance and perception) must be considered. Though aircraft perception is frequently evaluated with simple level-based metrics, higher level cognitive processes such as classification, identification, and localization benefit from additional signal features. To explore other non-level metrics, a pair-wise similarity experiment with four categories of sounds (including non-aircraft, rotorcraft, jets, and propeller aircraft) was performed by a panel of participants. These similarity ratings were decomposed into a timbre space using a two dimensional multidimensional scaling (MDS) analysis. This approach provided groupings that served as the baseline data for machine learning models to extract dominant features required to categorize aircraft signals. To model human responses, a musical feature extraction toolbox was used to obtain a set of audio attributes from the signals. These features were passed through a random decision forest, coupled with recursive feature elimination. The strength of this model is that it uses a data-driven approach to determine the appropriate mixture of waveform (or envelope) and spectral attributes that best capture participants' judgments. The model was validated by introducing a set of signals not presented to the participants, which were successfully placed in the timbre space corresponding to the correct category.