Abstract
A commonly used definition of spatial disorientation (SD) in aviation is ''an erroneous sense of one's position and motion relative to the plane of the earth's surface''. There exists a wide range of SD use-cases dictated by situational factors, therefore SD has been predominantly studied using reduced motion detection experimental contexts in isolation. The study of SD by use-case makes it difficult to understand general SD occurrence and thus provide viable solutions. To investigate SD in a generalized manner, a twopart Human Activity Recognition (HAR) study was performed. In Part I, a generalized SD perception dataset was created using whole-body experimental motion detection methods in a naturalistic flight context; joystick response was measured during rotational or translational vestibular stimulation. Results showed that SD occurred less for faster speeds than slower speeds, and specific orientations and axes were more difficult to detect motion. Part II evaluated supervised and unsupervised model parameters, including: model architecture, data use-case, feature-type, feature quantity, ground-truth labeling, unsupervised labeling. Long-Short Term Memory (LSTM), Random Forest (RF), and Transformer Encoder models most accurately predicted SD with mean accuracy of 0.84, 0.82, and 0.77 respectively. Using permutation importance (PIM), a dependency score for time, frequency, and time & frequency feature-types quantified the amount that each model architecture depended on a feature-type. The lenient ground-truth label best characterized features, and K-medoids clustering using position and velocity features most accurately replicated ground-truth labels.