Page 262 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

254 CHAPTER 12 Computational Intelligence in the Time

• Addressing subtle changes. Design of learning-based methodologies to detect
and anticipate subtle drift changes are needed, for example, to deal with aging at
the sensor and actuator level. In fact, slowly developing changes are hard to
detect, in the sense their magnitude is small and can be detected by current
change detection tests only in the long time period. However, it is expected that
availability of datastreams should permit to run machine learning tools to
estimate the current behavior of the features and build predictive models to
assess the level of time variance.

5. MODEL-FREE FAULT DIAGNOSIS SYSTEMS
Cyber-physical applications are mostly data-eager, in the sense that application
decisions and behaviors are strongly driven by the information content extracted
from a generally large platform of sensors. This dependency on sensor data can
be however very critical, since sensors and real apparatus are prone to faults and
malfunctioning that, in turn, negatively affect the information content carried by
data and used by the application to make decisions [15,16]. The problem amplifies
when low-cost sensors are considered and/or sensors are deployed in harsh and
challenging environments (e.g., think of a body network where sensors, external
buses, and connectors are subject to mechanical stress and environmental
challenges). As such, faulty sensors detection (also including sensors working in
suboptimal conditions) and mitigation are intelligent functionalities that must be
included in the design of any CPS to prevent propagation of erroneous information
to the decisional level [17]. At the same time, we comment that a generic method
designed to detect a fault occurrence will also detect any deviation in the information
carried by data, for example, caused by changes in the environment the sensor is
deployed in (time variance), and react erroneously: a novel family of methods are
hence requested to detect changes in the information content carried by data;
disambiguate between faults and violation of the time-invariant hypothesis; as
well as identify, isolate, and possibly mitigate the occurrence of faults. These tasks
are carried out by fault diagnosis systems (FDSs) (Fig. 12.2).

5.1 MODEL-FREE FAULT DIAGNOSIS SYSTEMS
Since CPSs rely on a rich and diversified set of sensors produced by a plethora of
companies, it is not possible to request an accurate physical model describing their
modus operandi. In this direction, model-free FDSs are requested to detect, identify,
and isolate the occurrence of faults without assuming that their signatures, the nature
of uncertainty, and their ruling equations are available.
Research on standard (not model-free) FDSs has provided major breakthroughs
in past decades by yielding several methodologies for detecting and diagnosing
faults in several real-world applications, for example, see Refs. [17,18]. However,
the effectiveness of a traditional FDS is directly proportional to the available