Towards R2D2: Real-time Identification of Time-varying Controlled Element Dynamics

Abstract

Enabling real-time identification and detection of changing pilot and controlled element (CE) dynamics in manual control tasks is essential for the development of real-time adaptive support systems (i.e., a “R2D2”) that 1) monitor closed loop dynamics and 2) detect irregularities therein. This study developed a method for identifying CE dynamics in compensatory target-tracking and disturbance-rejection tasks with time-varying pilot and CE dynamics. The CE dynamics were modelled using autoregressive exogenous (ARX) models, with the model parameters estimated using Ordinary Least Squares (OLS) and Recursive Least Squares (RLS). The effects of disturbance power, pilot remnant intensity, and RLS memory horizon on estimation bias and quality fit are studied using Monte Carlo simulations with 100 realizations. The simulation results are verified against the calculated least-squares best-fit model. A memory horizon of 512 samples (5.12 s at 100 Hz) was found to be optimal for all simulated configurations. The results show that for pure target-tracking tasks, a properly selected ARX model structure and an appropriately chosen RLS memory horizon allow for perfect model identification. In contrast, for pure disturbance-rejection, the identification results show the opposite. Increasing pilot remnant intensity was found to have a positive effect on both the mean parameter biases and quality of fit. With this work, an in-simulation functional approach for CE dynamics identification is established, together with a procedure for obtaining the least-squares best possible model fit for the CE ARX model, thereby providing a reference for the development and verification of real-time CE identification methods that, alongside real-time pilot dynamics identification methods, are necessary for the development of the monitoring functionality envisioned for R2D2-like systems.