Statistical Prediction of Failures in Aircraft Collision Avoidance Systems

Abstract

ACAS X is the next generation onboard collision avoidance system aimed at replacing the current standard TCAS for commercial aircraft. On-board collision avoidance systems are designed to help avoid dangerous Near Mid-Air Collision (NMAC) scenarios. Despite the fact that such systems can be very efficient in doing so, NMACs may still occur under rare circumstances. In this paper, we study the high dimensional time-series state space for encounters of aircraft equipped with ACAS X. We describe statistical modeling and learning techniques for predicting whether and when NMAC situations may occur. An iterative variable selection algorithm identifies the most influential variables for NMAC attribution. We also present a methodology for finding safety-boundaries, characterized as geometrical objects, that separate safe operational regions from dangerous ones where NMACs can occur. Even though our approach is presented in the context of ACAS X, it can be easily extended to numerous other domains including robotics, autonomous spacecraft, or self-driving cars.

Nomenclature

 

AC:

Aicraft

ACAS-X:

AC Collision Avoidance System

ALC:

Active Learning Cohn

ALM:

Active Learning MacKay

$\varDelta ^i_{alt}$ :

difference in altitude $i=1,2$

$\varDelta _z$ :

absolute vertical distance between AC

DynaTree:

Dynamic Regression Tree

E[...]:

Expectation

F1:

weighted average of P and R

FAA:

Federal Aviation Authority

FN:

False Negative

FP:

False Positive

I(X):

Improvement

LHS:

Least Horizontal Separation

NMAC:

Near Mid-Air Collision

P:

Precision

R:

Recall

$r^i_{target}$ :

target range AC$_i$, $i=1,2$

s:

slant range

SVM:

Support Vector Machine

$T_{NMAC}$ :

Time to NMAC event

TCAS:

Traffic Collision Avoidance System

TN:

True Negative

TP:

True Positive

UAV:

Unmanned Aerial Vehicle

$v^i_{vert}$ :

vertical speed for AC$_i$, $i=1,2$