Knowledge base: Warsaw University of Technology

Settings and your account

Back

Methodology of analysis of results of fault simulations using knowledge discovery methods

Agnieszka Komorowska

Abstract

Fault simulators are one of the key tools used to examine dependability of computer systems. The results of fault simulations are huge datasets which are analized in order to discover features of the examined software with significant impact on its reliability. Despite wide usage of knowledge discovery methods in other fields where huge datasets are analyzed, in fault simulations still the most popular tools are statistical methods. The purpose of this thesis is to adapt knowledge discovery methods to analyze results of fault simulations. The thesis proposes an universal methodology of discovering knowledge from data from fault simulators. During analysis of the raw data it was determined that these data have some specific properties. The most important is that they are multidimensional and multilevel. Based on this knowledge the author performed the critical analysis of pre-processing and model construction algorithms in relevance to the fault simulations data with this features. As the result of this analysis, multidimensional multilevel decision rules were chosen as the final data mining model. Some modifications of the Apriori algorithm of construction of decision rules were also proposed in the thesis. These modifications made possible: incorporation of information about attributes’ hierarchies into rules’ discovery process, filtering redundant patterns and visualisations of rules. Model evaluation criteria were also presented. One of the main parts of the methodology is evaluation of interestingness of rules. A selection of criteria of an interestingness measures evaluation suitable for rules build based on data from fault injections was made in the thesis. A method of evaluating rules’ interestingness within their neighbourhoods is also worth noting. The method consists of: a multidimensional multilevel rules’ distance measure, neighbourhood with variable radius definition and rule’s interestingness within its neighbourhood definitions. The new definition of rule’s neighbourhood with variable radius is adjusted to characteristic of multidimensional multilevel rules. Finally the paper describes a prototype of a specialised application to analysis data from fault injection experiments which implements the proposed methodology. All experiments presenting usage of the methodology were conducted using this tool. The methodology was applied to data from two fault simulators: FITS and QEFI which have different purpose and format of results. These experiments demonstrated effectiveness and generality of the proposed methodology.
Record ID
WUTe239afbf758d47d5b04225743355a63b
Diploma type
Doctor of Philosophy
Author
Agnieszka Komorowska Agnieszka Komorowska,, The Institute of Computer Science (FEIT/ICS)Faculty of Electronics and Information Technology (FEIT)
Title in Polish
Metodyka analizy wyników symulacji błędów z wykorzystaniem algorytmów odkrywania wiedzy
Title in English
Methodology of analysis of results of fault simulations using knowledge discovery methods
Language
(pl) Polish
Certifying Unit
Faculty of Electronics and Information Technology (FEIT)
Discipline
information science / (technology domain) / (technological sciences)
Status
Finished
Start date
25-09-2012
Defense Date
31-03-2015
Title date
21-04-2015
Supervisor
Internal reviewers
Tadeusz Łuba Tadeusz Łuba,, The Institute of Telecommunications (FEIT)Faculty of Electronics and Information Technology (FEIT)
External reviewers
Krzysztof Sapiecha Krzysztof Sapiecha,, Undefined Affiliation
Pages
177
Keywords in English
fault injection, software dependability, knowledge discovery, multidimensional
Abstract in English
Fault simulators are one of the key tools used to examine dependability of computer systems. The results of fault simulations are huge datasets which are analized in order to discover features of the examined software with significant impact on its reliability. Despite wide usage of knowledge discovery methods in other fields where huge datasets are analyzed, in fault simulations still the most popular tools are statistical methods. The purpose of this thesis is to adapt knowledge discovery methods to analyze results of fault simulations. The thesis proposes an universal methodology of discovering knowledge from data from fault simulators. During analysis of the raw data it was determined that these data have some specific properties. The most important is that they are multidimensional and multilevel. Based on this knowledge the author performed the critical analysis of pre-processing and model construction algorithms in relevance to the fault simulations data with this features. As the result of this analysis, multidimensional multilevel decision rules were chosen as the final data mining model. Some modifications of the Apriori algorithm of construction of decision rules were also proposed in the thesis. These modifications made possible: incorporation of information about attributes’ hierarchies into rules’ discovery process, filtering redundant patterns and visualisations of rules. Model evaluation criteria were also presented. One of the main parts of the methodology is evaluation of interestingness of rules. A selection of criteria of an interestingness measures evaluation suitable for rules build based on data from fault injections was made in the thesis. A method of evaluating rules’ interestingness within their neighbourhoods is also worth noting. The method consists of: a multidimensional multilevel rules’ distance measure, neighbourhood with variable radius definition and rule’s interestingness within its neighbourhood definitions. The new definition of rule’s neighbourhood with variable radius is adjusted to characteristic of multidimensional multilevel rules. Finally the paper describes a prototype of a specialised application to analysis data from fault injection experiments which implements the proposed methodology. All experiments presenting usage of the methodology were conducted using this tool. The methodology was applied to data from two fault simulators: FITS and QEFI which have different purpose and format of results. These experiments demonstrated effectiveness and generality of the proposed methodology.
PKT classification
4100
KBN classification
28 Informatyka
EU classification
8030
Thesis file
  • File: 1
    komorowska.pdf
    Request has been registered

Uniform Resource Identifier
https://repo.pw.edu.pl/info/phd/WUTe239afbf758d47d5b04225743355a63b/
URN
urn:pw-repo:WUTe239afbf758d47d5b04225743355a63b

Confirmation
Are you sure?
Report incorrect data on this page