Whereas a lot of research is focusing on ageing detection and
end of life prediction the problem of detecting safety critical
states using field data still lacks profound knowledge. In state-of-
the-art battery management systems (BMS) the compliance of
limits for critical signals such as voltage, current or temperatures
is supervised. Furthermore, the State of health (SOH) is monitored
to keep track of long-term ageing processes. But there is no
detection of anomalous states which may lead to catastrophic
failures.
Thus, identifying these anomalies in the battery systems during
its operation can increase its safety and viability. Identifying and
isolating these anomalous sections from the recorded battery
time series also helps to further analyze the anomalous
components later.
In the presented work, a novel probabilistic data-driven approach
to model the behaviour of the battery is explored. Various battery
signals, typically available within BMS, are analyzed. Thus, no
additional hardware or sensors are required. Beside system level
signals of current, voltage, temperature and soc also min and
max cell values of temperature and voltage are considered.
Thus also anomalous cell inhomogeneities can be detected.