The exploratory analysis of reliability data
The thesis outlines the usual parametric analysis of field failure time
data for repairable equipments. Due to shortcomings of this black-box
approach, exploratory reliability analysis has been adopted to exploit the
available data and so learn more about the physical failure process. Elements
of exploratory analysis have appeared in recent statistical applications of
point process, time series and multivariate methods in the area. These
approaches are reviewed and investigated.
Exploratory analysis of much field time between failure and limited
repair time data for hardware equipments has been undertaken. Despite
being from different physical mechanisms, software failure interval data has
the same underlying statistical point process as such hardware data and has
been similarly investigated.
Simple graphs, often with simulation bounds, inference procedures for
nonhomogeneous Poisson processes and Box-Jenkins analysis have been used to
search for and model aspects of structure expected in reliability data. The
appropriateness of the methods is discussed. As well as revealing that
(constant) failure rates are often unsuitable summaries, exploratory analysis
has highlighted features previously unknown or ignored. The identified time
structures, data irregularities and other complexities are described.
Exploratory analysis indicated potential dependent failures. A
simulation-based graphical tool for highlighting these important events is
described. Applications to real data have shown this is a promising approach.
Principal coordinates and cluster analyses have been used to explore
multivariate field data for automatic fire detection systems in an attempt to
identify circumstances leading to false alarms. Data problems limited this
Exploratory analysis has revealed it is common in reliability to assume a
too simplistic model formulation compared with the true complex data
structures. The implications of this for reliability data collection. storage and
analysis are discussed. While an exploratory approach is generally successful,
some specialisation of standard statistical methods for reliability is desirable.