Evolving Neural Network Weights for Time-Series Prediction of General Aviation Flight Data

Abstract

This work provides an extensive analysis of flight parameter estimation using various neural networks trained by differential evolution, consisting of 12,000 parallel optimization runs. The neural networks were trained on data recorded during student flights stored in the National General Aviation Flight Database (NGAFID), and as such consist of noisy, realistic general aviation flight data. Our results show that while backpropagation via gradient and conjugate gradient descent is insufficient to train the neural networks, differential evolution can provide strong predictors of certain flight parameters (10% over a baseline prediction for airspeed and 70% for altitude), given the four input parameters of airspeed, altitude, pitch and roll. Mean average error ranged between 0.08% for altitude to 2% for roll. Cross validation of the best neural networks indicate that the trained neural networks have predictive power. Further, they have potential to act as overall descriptors of the flights and can potentially be used to detect anomalous flights, even determining which flight parameters are causing the anomaly. These initial results provide a step towards providing effective predictors of general aviation flight behavior, which can be used to develop warning and predictive maintenance systems, reducing accident rates and saving lives.