NASA’s neural network project passes milestone

NASA researchers have completed a series of evaluation flights for a flight control system that could enable future aircraft suffering major system failures to be flown to a safe, controlled landing.

Researchers at NASA’s Dryden Flight Research Center, Edwards, California, have completed a milestone series of evaluation flights for a flight control system that could enable future aircraft suffering major system failures or combat damage to be flown to a safe, controlled landing.

The Intelligent Flight Control System (IFCS) research, on board a highly modified NASA F-15B aircraft, focuses on development of ‘self-learning’ neural network software for aircraft flight control computers. In its final form, the software would compare data from how the aircraft and its systems are operating with a database of how it would normally operate, and automatically adjust the flight controls to compensate for any damaged or inoperative control surfaces or systems.

The IFCS project team is said to have successfully met research objectives by evaluating in flight a passive online Parameter Identification (PID) algorithm (software code), and an online learning Dynamic Cell Structure (DCS) neural network algorithm.

According to NASA, this is a significant step for real time PID and neural net technology, and serves as a significant proof of technology for the project’s direct adaptive (Generation I) flight control concept.

The NASA team assessed the ability of the PID and DCS algorithms to efficiently identify aircraft stability and control characteristics, and map and retain this information as a function of flight condition.

The PID algorithm is an on-line function that determines the actual stability and control characteristics of the aircraft as it flies. When results from the PID algorithm differ from what is called the pre-trained neural network (PTNN), an update to the system is required.

The DCS provides the online learning of the system. It tracks the differences between the PTNN and PID and provides an organised map of updates to the stability and control derivatives of the aircraft.

The original DCS Code was delivered from NASA’s Ames Research Center, Moffett Field, California, then modified at NASA Dryden for Generation I application.

Included among important features of the DCS are the facts that it has long-term memory, critical for IFCS use, and has the ability to be enlarged by the addition of nodes.

IFCS software evaluations performed by the F-15B aircraft included handling qualities manoeuvres, envelope boundary manoeuvres, control surface excitations for real-time PID to include pitch, roll, and yaw doublets, and neural network performance.