AI for aviation

Pingbo Tang is creating AI tools that will one day aid workers in critical roles like aircraft control to avert risk and safely manage a complex system of moving aircraft. In his newest research, Tang, an associate professor of civil and environmental engineering, lays the foundation for future innovations in AI by identifying past factors that resulted in the loss of separation and then training a machine to recognize the factors associated with this risk to predict how better routing could prevent these scenarios.

Loss of separation is when two aircraft become too close, putting both at higher risk of a costly or potentially dangerous collision. Tang’s algorithm uses data taken via cameras monitoring the facial features of air traffic controllers, tracking and logging changes that correspond with a loss or near loss of separation incident.

Tang is working toward realizing several AI techniques for supporting the predictive management of facilities, workspaces, and civil infrastructure, all of which require algorithms like this that can quickly and efficiently process large datasets while overcoming missing data problems. Transfer learning is a machine learning technique in which researchers train a machine on large amounts of data with stored knowledge from similar problems and existing expertise on the subject. Such approaches can adapt models trained in some scenarios with sufficient training samples to other scenarios where data is limited.

Reinforcement learning is a machine learning technique involving training on large historical datasets. However, the machine is provided with different stored knowledge. It must produce a solution based on the analyzed data and clearly defined performance indicators of operational processes (penalties and awards that guide the machine to find the best strategies). Given the increased demands on a machine this would require, Tang says reinforcement learning is less tangible in the immediate future than a technique like transfer learning, which is informed by human expertise.

A tool based on reinforcement learning would be able to analyze datasets of aircraft positioning info—like those Tang used to train his map-matching algorithm—predicting and informing controllers how to avoid any taxi routes that would create a risk for loss of separation between aircraft, purely based on positioning data.

A tool based on transfer learning may use similar modeling and analytics. However, it would also be informed by information tagged by an expert and informed by stored knowledge from similar problems, such as his work on safety in nuclear power systems. Aircraft control involves a spatial system, while nuclear power plants are mechanical systems, but researchers like Tang can still draw many useful parallels for algorithms.