Page 263 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

5. Model-Free Fault Diagnosis Systems 255

FIGURE 12.2
A full fault diagnosis system is characterized by four phases: fault detection aiming at
identifying the presence of a fault; fault identification characterizing the type and nature of
the fault; fault isolation, localizing the fault; and fault mitigation, whose goal is to reduce
the impact of the fault on the system.

information and priors about the given system, in the sense that availability of (1) the
equations ruling the interaction between the cyber system and the physical phenom-
enon, (2) information about the inputs and the system noise, and (3) availability of
the “fault dictionary” containing the characterizations of feasible faults permits the
FDS to operate in optimal conditions. Even though some information might become
available in some very specific applications in general, it is hardly usable in
cyber-physical applications since the characterization of the CPS interaction, the
nature of existing noise and uncertainty, and the missing signature of expected faults.
Moreover, we cannot spend huge efforts in designing an FDS for any CPS-based
application if the requested procedure is too complex and articulated since costs
and time to market are fundamental requirements in designing successful
applications. Instead, we would appreciate a methodology able to automatically
learn the FDS directly from the data the application receives (computational
intelligenceebased model-free approach: no available model for the system under
investigation, no fault dictionary or fault signatures, no information about the nature
of uncertainty). In other terms, all unknown needed entities are learned from
available data through computational intelligence and machine learning techniques.
The particular computational intelligence technique depends on the information
available to solve a specific subproblem. For instance, we can use machine learning
and fuzzy systems to detect faults; the same technologies can be used to design a
fault dictionary and identify the type of fault through a classifier and find its
magnitude through inference. Fault localization can take advantage of a statisti-
cal/probabilistic or fuzzy logic framework, whereas mitigation can be based on
machine learning and fuzzy systems.
Existing model-free FDSs automatically learn the nominal and the faulty states
from sensor datastreams and take advantage of existing temporal and spatial rela-
tionships among sensors to detect a possible occurrence of faults. The learned