Taxiway Conformance Monitoring via ATC Voice Communications

Runway incursions are a growing concern in the United States (US) National Airspace System (NAS) as well as worldwide. The recent increased frequency of serious events at major US airports has prompted a focus on technologies capable of mitigating runway incursions by the Federal Aviation Agency (FAA) and other organizations. In 2024 alone, there were 1,757 runway incursions. Since 2008, pilot deviations have caused 63% of runway incursions in the US, many of which result from mistaken or misunderstood clearances by Air Traffic Control (ATC). In order to provide an early warning to controllers in the event of a pilot taxi deviation, MIT Lincoln Laboratory has developed a baseline taxiway conformance monitoring system capable of fusing air traffic voice communications with surface surveillance data to issue alerts when aircraft do not follow ATC clearances. To develop this system, we leveraged the closed-source Linguistic Data Consortium (LDC) ATC Complete dataset and an inhouse dataset containing time-aligned ASDE-X surveillance data and air traffic communication audio collected through the Runway Status Lights (RWSL) program. The alerting performance was evaluated on one hour of human-annotated RWSL data from Boston Logan International Airport. This paper documents the automatic speech recognition (ASR), natural language understanding (NLU), surveillance data processing, and developmental conformance alerting logic techniques. The ASR segment will outline baseline and fine-tuned performance of the NeMo Parakeet, Whisper, and Wav2Vec2 models on the LDC ATC dataset. It will also outline how models fine-tuned on the LDC ATC dataset perform on the inhouse RWSL dataset. The NLU section will demonstrate rule-based methods for the detection and parsing of nested entities of interest within the transcribed text, as well as boosted performance by incorporating airport and surveillance metadata. The surveillance data processing section will detail algorithms for identifying the taxiway or surface an aircraft is on. Finally, the conformance monitoring section will outline the alerting logic used to relate clearances extracted from ATC utterances to surface surveillance and issue alerts in the event of a mismatch. The end-to-end performance of the system on the one hour of RWSL data will be investigated and key areas for future improvement will be explored.