Convolutional Neural Network for Detection and Quantification of Pilot-Induced Oscillations

Abstract

Pilot-induced oscillations (PIO) are a long-standing challenge in aircraft handling, characterized by destabilizing oscillatory behaviour resulting from closed-loop interactions between the pilot and the aircraft. Traditionally, PIO identification and quantification rely on the subjective PIO Rating Scale (PIOR), used by pilots during handling qualities evaluations in aircraft testing and certification. This paper presents a data-driven approach to objectively identify PIO using flight data collected from a fixed-base simulator during a flight test campaign. The data were labelled according to the PIOR scale, preprocessed using the Wavelet Transform, and used to train a Convolutional Neural Network (CNN). This approach enables objective detection and quantification of PIO while maintaining alignment with the pilot-assessed evaluative framework. A k-fold cross-validation strategy yielded an average validation accuracy of 52.8\%, with the best-performing model achieving 63.1\%. The reduced overall accuracy is attributed to challenges in classifying low-to-intermediate PIO ratings (PIOR 2 and 3). Despite this limitation, the proposed model shows potential for further improvement and demonstrates promise as a complementary tool in handling qualities evaluations, providing a quantitative counterpart to traditional pilot assessments.