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.
1.2 Automated Fault Diagnosis Introduction<br />
Remote monitoring improves the diagnostic performance by providing a great amount of useful<br />
data. However, experts still have to interpret this data manually. Also knowledgable experts need<br />
to devote considerable time <strong>and</strong> effort to extract information <strong>and</strong> draw conclusions about faulty<br />
components. Automating this process is called automated fault diagnosis, <strong>and</strong> is discussed in the<br />
next section.<br />
1.2 Automated Fault Diagnosis<br />
Automated approaches provide many potential benefits compared to manual approaches. Automated<br />
fault diagnosis approaches automate the interpretation of raw data, in order to produce diagnoses.<br />
Such approaches are likely to be faster, less error prone, <strong>and</strong> less dependent on human<br />
intervention. Therefore, automating the fault diagnosis process can be one of the solutions towards<br />
improving dependability. Figure 1.1 shows how a diagnostic system could add to dependability. By<br />
processing data that is (remotely) monitored, the diagnostic system produces diagnoses <strong>and</strong> supports<br />
the service engineer in the diagnosing task. Then, the service engineer could focus on repair<br />
(the supervisory controller in the figure), <strong>and</strong> does not have to bother about the interpretation of<br />
log data. If the supervisory controller is also automated, a diagnosis that has been made within<br />
milliseconds could add to safety <strong>and</strong> reliability. The exact benefits depend on the specific approach<br />
<strong>and</strong> implementation of the diagnostic system. The remote monitoring technique produces data that<br />
contains information about faulty components. This is a practical environment to examine various<br />
automated approaches to fault diagnosis.<br />
The possible automated approaches can be characterized by black box <strong>and</strong> white box approaches.<br />
A black box approach uses externally observable behavior, but does not state anything about the<br />
structure of the system or behavior of internal components of the system. An example of a black<br />
box approach is a data mining technique that searches for correlations between log data <strong>and</strong> faulty<br />
components. It is unable to explain why a certain correlation exists. A white box approach uses<br />
internally non-observable behavior as well as externally observable behavior. The behavior of the<br />
whole system is defined by the structure <strong>and</strong> the behavior of internal components. An example of<br />
a white box approach to fault diagnosis are experts that manually define a mapping between symptoms<br />
<strong>and</strong> diagnoses. The mapping is implemented by application-specific code or, more generally,<br />
by using expert systems.<br />
Another characterization of approaches is whether the approach uses cause-to-effect or effectto-cause<br />
reasoning. The mapping of symptoms on diagnoses, as implemented by expert systems,<br />
is effect-to-cause reasoning. The next section introduces an approach that defines cause-to-effect<br />
relations.<br />
1.3 Model-Based Fault Diagnosis<br />
Model-Based fault Diagnosis (MBD) is a technique for doing fault diagnosis based on a model of<br />
a system. The model specifies all relevant information for doing fault diagnosis. A separate tool,<br />
a so-called diagnostic engine, operates on this model to pinpoint root causes of failures. It is first<br />
suggested by Reiter [25] <strong>and</strong> continued by de Kleer, Mackworth <strong>and</strong> Reiter [10].<br />
The idea behind the technique is shown in Figure 1.2. The cloud represents the real system,<br />
<strong>and</strong> its runtime operation. The model formally defines nominal, <strong>and</strong> possibly faulty behavior, of<br />
the system. The diagnostic engine uses this formal model for computing predictions of system<br />
behavior. During system operation, live data is gathered <strong>and</strong> processed in order to obtain actual<br />
4