I am looking into using a neural network to calculate the aerodynamic properties of aircraft. The following properties are required to calculate the aerodynamic properties of aircraft: aircraft geometry, aircraft speed, and aircraft altitude.
The aircraft geometry is represented by many 3D points, typically totaling 600 individual points/numbers for my application. Speed and altitude are represented by one number respectively.
The input file to the neural network would require the aircraft’s geometry, speed and altitude; and would therefore be very imbalanced: the geometry would have a much larger representation in the input file than the speed and altitude. This may create the problem where changes in speed or altitude aren’t picked up by the neural network because they represent such small changes in the input file. This is particularly bad because a small change in speed would change the aerodynamics greatly, whereas a small change in the geometry would not change the aerodynamics greatly.
Can a neural network deal with such large imbalances of importance in the input?
Alternatively, can I use convolutions to reduce the size of the geometry from 600 numbers to, for example, 5 to correct this imbalance? I could then combine this with the speed and altitude in a feed-forward neural network and have a much more balanced input.