pdf download - Software and Computer Technology - TU Delft
pdf download - Software and Computer Technology - TU Delft
pdf download - Software and Computer Technology - TU Delft
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
4.4 Modeling Model-Based Fault Diagnosis<br />
still be able to produce accurate diagnoses. Discretization is the process of mapping variables in a<br />
many-valued domain to variables in a few-valued domain.<br />
Some terminology regarding to this topic is useful. Note that this terminology is defined for this<br />
thesis, <strong>and</strong> might not correspond to terminology in other literature. Suppose there is a system that<br />
has 4 components, <strong>and</strong> logs 5 signals to a log file. Three of those signals are in the floating point<br />
domain (f1, f2 <strong>and</strong> f3), <strong>and</strong> two are in the boolean domain (b1, b2). Instances of the tuple (f1, f2, f3,<br />
b1, b2) are called system data, <strong>and</strong> this term is generally defined as:<br />
System Data: All unprocessed digital data that the target system produces.<br />
Each instance of the system data, for example (500, 400, 50, true, true), is part of a fault scenario,<br />
<strong>and</strong> the definition of this term is:<br />
Fault scenario: A particular failure of the system. It is characterized by the system data that is<br />
logged at the moment a failure occurs, in addition with the actual broken component.<br />
Often, the system data is hard to interpret, <strong>and</strong> its values should be discretized to a domain with few<br />
members. The term discretization is defined as follows:<br />
Discretization: A mapping of variables within the system data, that are in a many-valued domain,<br />
on observables, that are in a few-valued domain.<br />
Thus, discretization maps the variables in the system data to a domain with a finite number of<br />
members. It could be that the source domain of system data variables is unknown. Preferable, the<br />
few-valued domain has as few as possible members, say 2, 3, or 4. Each observable is a derivative<br />
of the system data. For example, suppose the system data (f1, f2, f3, b1, b2) is mapped on the two<br />
observables (o1, o2) by a discretization D. D defines two derivatives of the system data, say o1 =<br />
f1 - f2 - f3 > 10 <strong>and</strong> o2 = b1 ∨ b2. The result of the discretization for a fault scenario with values<br />
(500, 400, 50, true, true) is then (true, true).<br />
Because of the many-to-few mapping it is very likely that various fault scenarios are not different<br />
in the conclusions that could be drawn. These are said to correspond to the same fault category. A<br />
definition is:<br />
Fault category: A set of fault scenarios that have the same values for (a subset of) the observables.<br />
Suppose the system data of two fault scenarios is resp. (500, 400, 50, true, true) <strong>and</strong> (50, 20, 10,<br />
false, true). Discretization D maps the system data of both fault scenarios to the observables (o1,<br />
o2) = (true, true). Consequently, the MBD engine infers the same diagnoses. Therefore, these fault<br />
scenarios belong to the same fault category.<br />
The system data of the 3-inverter example, as well as the example of Section 4.6, only consists<br />
of variables that are already in the boolean domain. The next chapter presents a complex example<br />
that shows how a discretization simplifies modeling. It uses the concepts system data, discretization,<br />
fault scenario <strong>and</strong> fault category.<br />
4.4.4 Compositional Subsystems<br />
Another important issue is to make subsystem entities compositional. This is one of the characteristics<br />
of model-based computing. The 3-inverter example presented above is specified in a<br />
compositional model. A non-compositional model would be:<br />
42