applications.tex

We have evaluated the tools on a variety of ML-based systems.  In some cases, the system used ML to implement new aircraft functionality.  In other cases, the goal is to use a neural network (NN) to create a more time- or memory-efficient implementation of an existing function.  ML-based functions used for evaluation include:

\noindent{\bf Remaining Useful Life (RUL)~\cite{rul}}:  A convolutional neural network (CNN) uses vibration measurements from rotating equipment to estimate time until maintenance or replacement.  The RUL NN is reasonably large, with 1600 inputs (a sequence of snapshots of condition indicators from vibration sensors and other metrics) and 94,500 learnable parameters arranged in 12 layers. The output is the predicted remaining useful life of the equipment. A public version with training data and requirements to be verified was made available as a benchmark for the 2022 International Verification of Neural Networks Competition (VNN-COMP).


\noindent{\bf Recommended Cruise Level (RCL)}:  Computes time and fuel optimal altitude as a recommendation to the pilot, replacing a complex optimization calculation and saving CPU time. The NN is a fully-connected, feed-forward NN with rectified linear unit (ReLU) activation functions. It has 5 inputs (aircraft and environment conditions) and 2 outputs (time and fuel costs), and 5 hidden layers, each with 10 neurons.


\noindent{\bf Fuel Quantity Measurement (FQM)}:  Computes fuel mass based on sensor measurements, replacing less-accurate table-based implementations.  A NN is trained to invert a function that computes sensor measurements from fuel mass, fuel tank geometry, and aircraft orientation.  The NN is a fully-connected, feed-forward network with 6 sensor inputs and one output (fuel mass). The NN is very simple, having a single hidden layer consisting of 50 neurons and uses the \emph{tanh} activation function. 


\noindent{\bf Runway Overrun Protection (ROP)}:  Estimates aircraft landing distance based on weight, speed, weather conditions, runway slope, and other parameters according to the requirements of ED-250, Runway Overrun Awareness and Alerting System (ROAAS).  We evaluated one of the NNs from the system which was a fully-connected, feed-forward network with 10 inputs and 3 outputs.  The outputs correspond to stopping distances with different brake settings. The NN has 2 hidden layers with 40 neurons each. It uses the \emph{tanh} activation function. 

\noindent{\bf Flight Trajectory Optimization (FTO)}: A neural network-based implementation of an optimized trajectory function, such as the A* algorithm and derivatives, to reduce computation time in a set of complex flight and weather conditions. The NN is a fully-connected, feed-forward network with ReLU activation functions. It has 7 inputs, which model aircraft position relative to an assigned flight altitude limit, distance to weather/threat, and relative velocity. It has 3 hidden layers with 35, 70, and 70 neurons, respectively. It has 5 outputs representing the next flight direction: up, down, right, left, and straight.