Sensor Integration for Inflight Icing Characterization
Using Neural Networks
James W. Melody, Devesh Pokhariyal, Jason Merret,
Tamer Basar, William R. Perkins, Michael B. Bragg
University of Illinois, Urbana, Illinois 61801
ABSTRACT
This work advances a neural network that characterizes
aircraft ice accretion in order to improve flight performance and safety.
Neural networks have been developed previously for use within an ice management
system that monitors inflight aircraft icing and its effects upon performance,
stability, and control. The previous work has applied these networks to
stability and control derivative estimates provided by an H°° parameter
identification algorithm during a longitudinal maneuver. This paper extends
those results by addressing ice characterization in the absence of pilot
input when poor excitation of the flight dynamics limits the accuracy of
parameter estimates. To compensate for this shortcoming inherent to steady-level
flight scenarios, the neural network presented in this paper integrates
steady-state characterization and hinge moment sensing with parameter estimates.
The neural network provides icing characterization in terms of an estimate
of the previously developed icing severity factor, rj. Extensive simulation
results are presented that indicate the accuracy of neural network characterization
during steady-level flight in the presence of sensor noise and turbulence
over a broad range of flight trim conditions and turbulence levels. Furthermore,
the relative utility of each information source is investigated via consideration
of network accuracy of networks trained only on that information source.