25.07.2014 Views

pdf download - Software and Computer Technology - TU Delft

pdf download - Software and Computer Technology - TU Delft

pdf download - Software and Computer Technology - TU Delft

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Model-Based Fault Diagnosis<br />

4.4 Modeling<br />

The only difference is the way how relevant information is specified. As explained in Chapter<br />

3, this effects the time to create the model, is error-prone, <strong>and</strong> inflexible in case of a design<br />

change. But is some cases there is no other possibility.<br />

4.4.2 Granularity<br />

The first step of making a model is deciding the granularity, in other words ’level of detail’, of the<br />

model. Granularity has two dimensions; depth <strong>and</strong> breadth.<br />

If you consider depth in granularity, it is possible to tune the level of abstraction. For each<br />

system, there is an infinite number of levels of abstraction. Consider the 3-inverter example again.<br />

On the highest level of granularity the 3-inverter system has one entity; the 3-inverter system. In<br />

this case, the LYDIA model has one health variable, <strong>and</strong> the only outcome of the diagnostic engine is<br />

whether or not the system is broken. This means that, in case of malfunctioning, a supervisory controller<br />

would have to replace the entire system with a healthy one. In this case, the costs to recover<br />

the system are higher than when a model that specifies the system at a lower level of granularity is<br />

used. On one lower level, the model distinguishes 3 entities, namely the 3 inverters, as shown in Figure<br />

4.1. The results of these diagnoses state for each inverter if it has to be replaced with a healthy<br />

one, or that it could be preserved. Going another step lower could be that the model also specifies<br />

the transistor <strong>and</strong> resistor that (could) implement the inverter. But usually these components cannot<br />

be replaced, <strong>and</strong> diagnosing this level of detail is of no added value.<br />

If you consider breadth in granularity, it is possible to include or exclude certain components<br />

from being modeled. Given a certain level of granularity, what components should, <strong>and</strong> which<br />

should not be specified by the model? For example, the 3-inverter system might include a casing<br />

that protects the system from the outside. Should this component be included? Maybe the casing<br />

gets easily damaged, <strong>and</strong> is the cause for the breaking down of inverters in the first place. If the<br />

casing never breaks, adding it, would only unnecessarily increase complexity. The particular considerations<br />

greatly depend on the specific system, the environment, <strong>and</strong> the motives of the modeler.<br />

As said before in this thesis, modeling of a system is not a trivial activity. Although modeling<br />

is a divergent activity without general procedures, this thesis considers three possible approaches to<br />

modeling:<br />

1. Modeling the entire system. The modeler chooses its granularity, based on domain-dependent<br />

<strong>and</strong> general considerations (e.g. Could the supervisory controller use the correct or incorrect<br />

functioning of this component for recovering the system? Does this component ever brake<br />

down? Is it possible to, directly or indirectly, observe the state?)<br />

2. Modeling to cover faults. The modeler composes a list of all known faults (things that could<br />

brake down). Only those components that could reveal a fault are included in the model.<br />

3. Modeling based on log data. Most existing systems log values of certain parameters to allow<br />

for monitoring the state of the system. Experts, mostly developers of the system, know what<br />

to conclude about the system state given a set of parameters. The modeler chooses the granularity<br />

of the system in order to reproduce the same diagnosis for a certain set of parameters.<br />

Of course, a modeler could use more approaches in parallel, in order to achieve a high quality model.<br />

4.4.3 Discretization of Observations<br />

The output of sensors, or other sources of data in a system, usually contains much resolution. Most<br />

sensor output is continuous. It is hard to use variables with high levels of resolution in a model, <strong>and</strong><br />

41

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!