In the
context of H2020 project I-NERGY
[1], R&D Nester has been testing
novel techniques for fault analysis and condition based monitoring strategies,
using Circuit Breaker historical data.
One of the
tools developed is an advanced
methodology for Fault Detection, based on oscillography data. It proposes
the usage Change Point Detection (CPD) algorithms, together with a set of signal
processing and rule based decision processes, to implement a fully data-driven
Fault Detection module, able to filter out fault records from other events.
Pruned
Exact Linear Time (PELT) is an highly popular and efficient Dynamic Programming
algorithm for CPD, introduced by Killick [2]. While many CPD methods are either
fast but heuristic or exact but slow (e.g. Binary Partioning), pruned methods
look to be both fast and exact [3]. The common approach for CPD is given by the
optimization of the following problem:
Where
is the input signal,
is the vector of changepoints,
is the cost function for a segment and
is the penalty against overfitting. Most exact
approaches solve this problem using an iterative timestep approach, where
previous timesteps are considered candidates for the previous changepoint to
the current time [3]. Pruning increases the computational efficiency of this
Optimal Partitioning approach (performing in
assuming a linear distribution of changepoints),
while ensuring that a global minimum is found. This is done by pruning the
number of candidates at each iteration, assuming that when a changepoint is
introduced, the cost necessarily reduces.
In the context of this
work, PELT is applied to transformed versions of the analog channel data. Each
of the time-series is standardized by mean and variance, and an approximation
of the envelope is obtained by the application of the Hilbert Transform.
The system
is designed to classify each COMTRADE files into 3 distinct scenarios, using
the number of changepoints detected:
1. Normal operating conditions: triggered by software errors or misconfigured protection systems,
these do not
have any oscillographic event.
This scenario is associated to changepoint absence.
2. Fault: which
represents a typical shunt fault, such as phase to ground fault. Typical
scenarios include more than
one changepoint in any of the input signals.
3. Planned Circuit Breaker opening: a standard maneuver, for maintenance purposes.
The typical pattern, in these
cases, are single occurrences of changepoints
which are followed by absence of current.
The
implemented methodology is integrated in the toolbox developed for the I-NERGY
R&D NESTER pilot, where there historical event data must be sorted and
faults should be separated from other tripping events.
References:
[1] I-NERGY Project Consortium. I-NERGY - Artificial
Intelligence for Next Generation Energy. Online. Available at:
https://i-nergy.eu/.
[2] Killick, Rebecca & Fearnhead, Paul &
Eckley, I.A. (2012). Optimal Detection of Changepoints with a Linear
Computational Cost. Journal of the American Statistical Association.
[3] Maidstone, R., Hocking, T., Rigaill, G. et al.
(2017). On optimal multiple changepoint algo-rithms for large data. Stat Comput
27, 519-533. https://doi.org/10.1007/s11222-016-9636-3.