Abstract:

The growth in size and scope of data is far surpassing our capability to meaningfully extract value from it – Air Force pilot training data is a prime case study. The proliferation of modern simulators, distributed AR/VR training devices, aircraft with dedicated IT linkages, and vast data lakes is already underway. However, novel analytical methods are required to make sense of these data in a format and timeframe that can be used for active instructional feedback without incurring the significant upfront time / resource costs normally associated with data labeling, structuring, scrubbing, and aggregating.

This paper will extend a modular framework for responsive pilot training presented at I/ITSEC 2024 towards a specific capability gap: the extraction, identification, and assessment of flight maneuvers from unlabeled time series data. Flight training systems generate enormous amounts of data, often in standardized formats from the simulation or gaming community. These datasets capture time, space, and position information (TSPI) of different event actors in sufficient detail to play back the event for instructional purposes. However, these standards do not include any form of analytics on the actions, how they relate to the context of the event, or how they should be judged against other students / training standards.

A novel application of a computationally-efficient, time-series pattern matching algorithm has been applied to multiple student flight regimes, and results will be discussed. As opposed to other methods that may require large, labelled datasets or extensive domain knowledge to define rules-based engines, the Matrix Profile algorithm can be applied directly to time series data generated by simulator or aircraft with minimal setup. Applications of this approach to active instructional querying of event logs and automated maneuver performance assessment will illustrate the tangible benefits to military flight training.

Keywords: ANALYTICS;LEARNING ANALYTICS;PERFORMANCE