Burn-in is the preconditioning of assemblies and the accelerated power
-on tests performed on equipment subject to temperature, vibration, vo
ltage, radiation, load, corrosion, and humidity. Burn-in techniques ar
e widely applied to integrated circuits (IC) to enhance the component
and system reliability. However, reliability prediction by burn-in at
the component level, such as the one using the military (e.g., MIL-STD
-280A, 756B, 217E [23-25]) and the industrial standards (e.g., the JED
EC standards), is usually not consistent with the field observations.
Here, we propose system burn-in, which can remove many of the residual
defects left from component and subsystem burn-in (Chien and Kuo [6])
. A nonparametric model is considered because I)the system configurati
on is usually very complicated, 2) the components in the system have d
ifferent failure mechanisms, and 3) there is no good model for modelin
g incompatibility among components and subsystems (Chien and Kuo [5];
Kuo [16]). Since the cost of testing a system is high and, thus, only
small samples are available, a Bayesian nonparametric approach is prop
osed to determine the system burn-in time. A case study using the prop
osed approach on MCM ASIC's shows that our model can be applied in the
cases where 1) the tests and the samples are expensive, and 2) the re
cords of previous generation of the products can provide information o
n the failure rate of the system under investigation. (C) 1997 John Wi
ley & Sons, Inc.