Bayesian Change Point Detection for Satellite Fault Prediction
In: Interdisciplinary Graduate Conference (IGC) 2010, 28-29 June 2010, Cambridge, UK.
Change point detection makes predictions in data whose structure changes
over time and belongs to a class of methods called time series methods.
Machine learning is the design of algorithms whose performance improves
with data. Techniques from statistics, such as change point detection,
have received increased interest in recent years in machine learning.
Many problems have change point structure: Stock markets exhibit
change points when some significant economic event causes an increase
in volatility, weather systems may exhibit change points during years
of El Nino, and electronic systems show change points when devices
begin to fail or configurations change.
This paper focuses on detecting changes in the operation of satellite
base stations. The aim is to provide predictions of device lifetimes
or time until maintenance will be necessary. Typical reliability
engineering uses survival analysis to find lifetime distributions
and mean time between failures (MTBF). However, using change point
detection we can make predictions that include information dynamically.
For instance, a change in the measured weather conditions could signal
an impending satellite signal loss due to an imminent storm. Modeling
changing dependencies is also important; a change in the relationship
between the external and internal temperature on a device should
mean there is a problem with the cooling unit. In the mentioned examples,
the weather change point occurred when the weather conditions changed
and when the cooling unit began to malfunction, respectively. Unlike
MTBF, change point detection provides probability distributions describing
the likelihood of a failure, which is necessary in most cases since
there is usually significant remaining uncertainty over the time
until failure. The two key aspects are that change point detection
can include information dynamically and model changing dependencies.
We show how to use techniques from Bayesian statistics, namely change
point detection, to model and understand electronic systems.