IEEE - Aerospace and Electronic Systems - July 2021 - 84

Machine Learning/Artificial Intelligence for Sensor Data Fusion-Opportunities and Challenges
Table 3.
Data-Driven Versus Model-Based Techniques Concerns
Data driven: (DL
methods).
Figure 4.
AI/ML integration for SDF.
PRAGMATIC DEPLOYMENT ISSUES FOR AI/ML
Wolfgang Koch highlights that the tenets of AI should
benefit society. For example, the wheel was not seen in
nature, but has profoundly changed human society. Hence,
artificially intelligent and technically autonomous systems
assist pragmatic perception and action. The AI and SDF
communities have many things in common as humans
need to perceive and act; however, AI comprises much
more than the recent progress in NNs and DL.
Using the fusion community DFIG model, then there
are two types of issues: (1) data-driven: low-level data
fusion for object assessment and (2) model-based: understand
causality for higher level data fusion of situation
assessment [10]. NN/DL utilizes signals and data to form
a model, but the model does not completely explain the
situation context as shown in Figure 4.
Figure 4 represents the DFIG fusion levels (see
Figure 2) in a manner that highlights AI (ML, DL, NNs),
impacting the core of SDF for assessment. The surrounding
elements include contextual influences (goals, environment)
feeding the sensor opportunities that are selected
by the operator (i.e., Level 5 user refinement), to manage
the resources (i.e., Level 4 process refinement).
The challenge of data-driven and model-based AI
fusion methods is " serious use. " For example, models need
to adhere to predictable and reproducible properties much
as the derivation of first-principle physics models. The construction
of AI models needs to be insensitive to various
" unknowns, " robust to noise, and free from attack. The AI
systems require adaptation to operational conditions, graceful
performance degradation, and methods of explainability.
The final aspect for AI adoption is certification such as
adherence to and compliance with a " code ofconduct. "
The critical questions are what and why of incorporating
AI/ML into data fusion systems as described in
Table 3. Data-driven methods can process vast amounts of
data with dedicated advances in hardware from which
there is a growing market. Data-driven methods with the
84
appropriate metadata can discern where and when (e.g.,
object assessment), but lack the ability to derive the why
and for whom (e.g., threat and impact assessment)-which
is an unsolved research challenge of explainable AI (xAI)
[11]. There are opportunities to resolve explainability by
combing AI/ML with SDF. To a large extent, the fusion
community has leveraged sensor management as a method
that predicts future states and then collects data to rule out
unlikely events through multiple hypothesis methods for
situation explainability.
The power ofDL is in the big data analysis towards ever
finer parameter estimation ofsalient variables. Using Bayesian
DL affords measures of reliability, proven probabilistic
methods, integration ofcontext, scalability, sparse data analysis,
and robust systems engineering. Examples of the benefits
of AI/ML and SDF include sequential Monte Carlo
methods for long short-term memory prediction in target
tracking [12]. Likewise, the ability to predict behavior adds
" induction " into " deductive " probabilistic inference.
One of the challenges for both AI/ML and SDF is context.
Context is an element of a situation from which
dynamic systems have to learn context while training and
in many cases for the unknown. Thus, there is a need for
human-support to infer the situation. A critical question
relates to the infamous philosophical qualia question [13],
which is the level of consciousness attended to by a user.
For example, the Qualia Exploitation of Sensor Technology
seeks to improve the decision quality ofa set ofagents
(human or machine) with three assumptions [14]:
Assumption 1: Fundamental units of conscious
cognition are situations.
Assumption 2: Decision quality is dominated by the
appropriate level ofsituational awareness.
Assumption 3: Cohesive narratives are reported
products of information fusion systems.
IEEE A&E SYSTEMS MAGAZINE
JULY 2021
Boosting: vast data
and hardware
Training: gather
enough data samples
Deception:
adversarial examples
Context: learn along
with data
Correlation: " Tell
me why? "
Model based: (Bayes
Reasoning methods)
Estimation:
" uncertain " logical
reasoning
Testing: systematic
algorithm design
Action: probable
cause-effect chains
Context: utilize expert
knowledge
Association: " Tell me
how? "

IEEE - Aerospace and Electronic Systems - July 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - July 2021