Each edge has an assigned weight, corresponding to the probability that “element a has an influence on element b.” Note that influence and dependence might
not be symmetrical.
Every element in the system is described by a pattern function that compares the
received (monitored) data with predicted data. During system operation when a
discrepancy is detected between real and the predicted (expected) data, within a
given threshold, the MASS performs the following steps:
– Using the dependency matrix, starting from the “suspected” element, all possible scenarios, i.e., the full graph of all dependencies the suspected element has,
are evaluated and their probability is calculated. Probabilities along the paths are
cumulatively multiplied. The resulting scenarios are then ranked according to
likeliness using a predefined threshold;
– The structure of the dependency matrix and all probabilities are assumed to be
constant during the graph unfolding;
– The resulting graph with the suspected root element is then treated as a fault tree
and evaluated either automatically or by a human to find the real reason for the
detected anomaly.
Hidden faults, faults that do not immediately result in errors but are triggered later
in time, must be given special attention. The introduced fault latency occurs
because the manifested discrepancy between the real and the expected outcome of
an element (i.e., program output) might be either wrong due to wrong behavior of a
suspected element itself or due to faulty elements the suspected element depends
on.
In other words, an element which shows faulty behavior is not necessarily faulty,
and the real source could also be another element that influences the first element.
We therefore introduce a search algorithm that recursively goes through all
dependencies starting with the suspected one.
The algorithm uses the assigned probabilities on the dependencies to calculate
the likeliness that an element is the source of the fault resulting in a list of possible
sources and their likelihood. Every element of this list is then checked to efficiently
find the real cause.
We propose to use the dependency matrix as data structure to store all element
dependencies. Fault tree schemes used for description of a system for diagnostic
purposes do not look like reasonable alternative to the dependency matrix for
several reasons:
– Fault trees need more space and more effort to create them manually.
– Fault tree is static by design and for each entry—new case of fault we have to
have new fault tree formation.
– Application of fault tree in real time of system operation is impossible.
We have proposed a dependency matrix with introduced GAFT. This matrix is then
used to derive the fault tree and all dependencies in real time [1].
Furthermore, MASS assumes concurrent processing of present and incoming
information about system elements to update the dependency matrix after each
48
5 GAFT Generalization: A Principle and Model of Active System…
not be symmetrical.
Every element in the system is described by a pattern function that compares the
received (monitored) data with predicted data. During system operation when a
discrepancy is detected between real and the predicted (expected) data, within a
given threshold, the MASS performs the following steps:
– Using the dependency matrix, starting from the “suspected” element, all possible scenarios, i.e., the full graph of all dependencies the suspected element has,
are evaluated and their probability is calculated. Probabilities along the paths are
cumulatively multiplied. The resulting scenarios are then ranked according to
likeliness using a predefined threshold;
– The structure of the dependency matrix and all probabilities are assumed to be
constant during the graph unfolding;
– The resulting graph with the suspected root element is then treated as a fault tree
and evaluated either automatically or by a human to find the real reason for the
detected anomaly.
Hidden faults, faults that do not immediately result in errors but are triggered later
in time, must be given special attention. The introduced fault latency occurs
because the manifested discrepancy between the real and the expected outcome of
an element (i.e., program output) might be either wrong due to wrong behavior of a
suspected element itself or due to faulty elements the suspected element depends
on.
In other words, an element which shows faulty behavior is not necessarily faulty,
and the real source could also be another element that influences the first element.
We therefore introduce a search algorithm that recursively goes through all
dependencies starting with the suspected one.
The algorithm uses the assigned probabilities on the dependencies to calculate
the likeliness that an element is the source of the fault resulting in a list of possible
sources and their likelihood. Every element of this list is then checked to efficiently
find the real cause.
We propose to use the dependency matrix as data structure to store all element
dependencies. Fault tree schemes used for description of a system for diagnostic
purposes do not look like reasonable alternative to the dependency matrix for
several reasons:
– Fault trees need more space and more effort to create them manually.
– Fault tree is static by design and for each entry—new case of fault we have to
have new fault tree formation.
– Application of fault tree in real time of system operation is impossible.
We have proposed a dependency matrix with introduced GAFT. This matrix is then
used to derive the fault tree and all dependencies in real time [1].
Furthermore, MASS assumes concurrent processing of present and incoming
information about system elements to update the dependency matrix after each
48
5 GAFT Generalization: A Principle and Model of Active System…