Aerospace and Electronic Systems Magazine April 2017 - 15
Imai et al.
Further research is needed in understanding
and formalizing the relationships between
hundreds of data streams that are available
to flight systems. In the Air France Flight
447 case study, the relationship between
speed vectors was mathematically simple
(i.e., Equation (1)), whereas in the Tuninter
1153 case study, the relationship between
weight and aircraft performance was not
so easy to formalize since different factors
beyond weight affect air speed. Figure 14
shows the relationships between the two
While it is possible to use aerodynamics theory to relate lift and weight Figure 14.
Examples of data stream relationships in AF447 and TU1153.
(equal in cruise phases of flight) to airspeed, other factors such as altitude,
temperature (affecting air density), angle of attack, and engine setfault-tolerance enabling safer flight. In this article, we overviewed
tings play an important role in the data relationship.
the ProgrammIng Language for spatiO-Temporal data Streaming
applications (PILOTS) system which is a declarative language.
With tens of lines of code, a data scientist can test and validate
their error detection models in the form of error signatures. Examples were presented to demonstrate dynamic data-driven error
Sensor fusion analysis for aircraft safety includes three areas: (1)
detection and correction using data from the Air France Flight 447
health monitoring, (2) pilot awareness, and (3) air-ground coordiand the Tuninter Flight 1153 accident reports. We view our work
nation. For health monitoring, there are many opportunities that inon error signatures as complimentary to existing work on fault declude the DDDAS paradigm with multimeasurement sensor fusion,
tection, isolation, and reconfiguration. In particular, it is possible
software enhancements, and uncertainty modeling . Using the
to use high-fidelity models that constrain the relationship between
normal flight operations, models can be used to cross-check sensensor and actuator data following aerodynamic principles. Such
sor readings and with our proposed software and techniques ,
systematic analysis of fault behaviors can result in a set of residualert when unexpected sensor readings occur. Inherently, sensor
als and error signatures that improve the performance and accuracy
fusion techniques can determine inconsistency between readings.
of fault detection, isolation, and reconfiguration strategies in flight
Examples of redundant sensor monitoring would best be achieved
systems. Such additional avionics support is needed to alert pilots
with more than two sensors to resolve conflicts between only two
with warnings of sensor faults in a more effective way for timely
sensors . The second category supports pilot awareness which
diagnostics and safer flight.
in and of itself is the leading cause of accidents (e.g., pilot error).
Future work includes exploring automated error signature
Sensor fusion would be able to use multiple sensors to display
acquisition by using machine learning and distributed computing
warnings in the cockpit. The challenge is to experiment (e.g., with a
, . Machine learning techniques  may help physicssimulator) the optimum display technology to warn pilots of critical
based models be more accurate by taking data into account. A dichallenges. There have been many accidents when the pilot failed
rection of future work for the PILOTS project is to be expanded
to take action even in the presence of stall lights and horns .
into a DDDAS Model Learning Toolkit: to facilitate Monte Carlo
The final category is air-ground coordination which could be both
simulations to learn model parameters from data, to enable Kala machine (e.g., GPS) and communications (e.g., with Air Traffic
man filters to reduce the impact of noise in the data, and to use
Controllers) that alert to low approaches, see-through bad weather
probabilistic (Bayesian) techniques to continuously update modconditions, relevant pilot reports and data from other aircraft, and
els to account for new data. Distributed computing helps address
identification of unavoidable terrain. The future of sensor fusion for
scalability as the number of data streams to be analyzed increases
aircraft accident mitigation is in a variety of methods to provide exwith the increased complexity of avionics systems. For example,
ternal validation of unsafe flight that can enhance action plans when
due to the increasing number of sensors in aircraft, more complex
health monitoring and pilot awareness are not sufficient.
aircraft fault models, error signatures, and damaged aircraft performance profiles would need to be assessed. Two methods are
CONCLUSION AND FUTURE WORK
being investigated: high-level abstractions and uncertainty quantification. High-level abstractions  enable data scientists and
Supplementing flight systems with error detection and data estimaengineers to more easily develop concurrent software to analyze
tion based on error signatures can add another layer of (logical)
IEEE A&E SYSTEMS MAGAZINE