Page 264 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

256 CHAPTER 12 Computational Intelligence in the Time

relationships are then used to characterize the system nominal state, as well as detect
any deviation from such a nominal behavior, diagnose the causes of the deviation,
identify the nature of the faults, isolate faulty sensors, anddpossiblydmitigate their
effects. It must be pointed out that most of existing solutions either apply the
learning mechanism only to a particular aspect of the FDSs (e.g., the fault
dictionary) or solve specific applications: very fewdpreliminarydmodel-free
methodologies have been proposed in the literature, for example, see Ref. [19].
Such solutions aim at characterizing the relationships present in the acquired
datastreams to autonomously learn the nominal state and construct, whenever
possible, the fault dictionary during the operational life of the system for fault
detection, isolation, and identification purposes.
However, despite these encouraging results major investigations must be
accomplished to reach the maturity level needed to support an automatic
design of a model-free FDS for networked cyber-physical applications that are
automatic, effective, control false positives/negatives, and are computationally
light.

5.2 RESEARCH CHALLENGES
In order to support the next generation of any cyber-physical application, we need to
address some open research issues
• Multiple faults. The “single fault” assumption has to be relaxed to host multiple
“concurrent” faults, possibly also of transient type, as it is the case in
cyber-physical/human applications. In fact, once a fault occurs it is likely that a
domino effect will arise and a subset of sensors is affected. Graph-based
machine learning techniques are expected to be the right tools to address
this research topic.
• Disambiguation module. Once a change in stationarity is detected, we need to
run powerful methods able to disambiguate among changes in the environment
and faults and false positives introduced by the change detection method. Given
the nature of the problem and the type of information available, it is expected
that machine learning and fuzzy tools should be the appropriate techniques to be
applied here.
• Unbalanced data. Faults are rare events; as such, it is hard to have many
data coming from the “faulty class.” This implies that we need to provide
machine learning methodologies to design effective FDSs starting from few and
unbalanced data.
• Modeling aspect. We need to provide novel design methodologies for model-free
FDSs. Such methodologies should be able to automatically configure the FDS
for the given application after having profiled sensor data. It is expected that
machine learning techniques should be the right tools to be used here to design
such a system.