Detecting Operator Distraction Through Multivariate Signal Processing

Abstract

This paper compares different multivariate signal processing methods - both based on classical statistical or machine learning models - for detecting human operator distraction in a preview tracking task. Binary Time Series Classification (TSC) methods proposed in prior work, which require both 'non-distracted' and 'distracted' training examples, were found to be out-performed by dedicated Anomaly Detection (AD) methods that aim to detect outliers from a learned 'non-distracted' prior distribution only. Even when applied in an ensemble learning framework, the TSC methods suffer from high false alarms and are highly sensitive to domain shift, due to their explicit reliance on representative 'distracted' training data. The considered statistical AD method, based on calculation of the Mahalanobis Distance (MD) of tested multivariate time-series samples, was found to be outperformed by a machine learning One-Class Support Vector Machine (OCSVM) method, as the latter more faithfully captured the (non-normal) distribution of the 'non-distracted' training data. Combining both AD methods in a meta detector only gave marginal improvement in the attained F1 score (0.5694) compared to the OCSVM (0.5610). The presented findings show that statistical and machine learning methods for AD can enable relevant detection of potentially impactful moments of distraction from human control data, which can be tested in real-time monitoring and adaptive support frameworks.