Page 13 - CE_Industral_Journal_Jan2008

A Novel Approach for Ground Fault Detection

The second neural network design used the spectrum Table 1. Change in Detection and False Alarm Rates
of the 3-cycle window of data. The magnitude of the with Threshold.
FFT of the 1000 samples was truncated at the 13th
harmonic. This resulted in a reduction to only 40 input
nodes for the neural network. This network had fewer
weights and biases and could be trained almost an or-
der of magnitude faster. The best results occurred when
30 nodes were used in the hidden layer. The network The results indicate that the network using the spec-
was trained with 600 cases and had a sum-squared trum (FFT) of the monitored current is more capable of
error of 11.8 (8 missed detections and 4 false alarms). detecting high impedance fault than the network using
Generalization testing on 3600 new inputs resulted in the actual current samples. Using the sampled current
about an 86.06% detection rate with about a 17.06% network in tandem with the spectrum based network
false alarm rate. The increased performance of this can reduce the false alarm rates, however, it doesn’t
network over the previous network is likely due to the appear to increase the detection rate significantly.
invariance of the frequency spectrum to phase shifts. The lack of synchronizing the current’s zero-crossing
These performance figures are once again based upon during training and generalization may prohibit this
using about 0.5 as the output threshold for indicat- neural network from detecting some of the patterns or
ing a presence of high impedance faults. An attempt features attributed to high impedance fault currents,
was made to reduce the false alarm rate by increas- such as asymmetry of half cycles and variations from
ing the output threshold to about 0.75. This resulted cycle to cycle. The results are encouraging given that
in about an 83.7% detection rate with about a 14.8% the detection is performed on only a 3 cycle snapshot
false alarm rate. Increasing the threshold to about 0.95 of data.
resulted in about a 77.7% detection rate and about an
11.8% false alarm rate.
Third neural network architecture was a combination Safe.
of the two previous networks operating in parallel. If
the output of both networks was greater than 0.5, then Economical.
a positive detection decision was indicated. For the
cases in which the two neural networks disagreed as to
the presence of a high impedance ground fault cur- Reliable.
rent, the output of the two networks was summed and
a variable threshold was used to make the decision. A
threshold of 1.0 corresponded to making the final deci- • Enhanced pole inspection with StrengthCalc ™
sion based upon which network was more confident in and LoadCalc ™
its own decision. • Lower replacement costs with the C2-Truss ™
For example, if the output of network 1 was 0.9867
and the output of network 2 was 0.0175, then the • Accurate data for GIS, OMS and joint use with
Video Data Capture
sum would be less than 1.0 and no decision would be
made because the output of network 2 is closer to the • Web-based records with Osmose
ideal value of 0 than the output of network 1 is to the Mapping Center
ideal value of 1. • Storm Response and Recovery
On the other hand, if a more conservative approach
were desired in which one chose to reduce the false Your Local Caribbean Osmose Professional is:
alarm rate, a larger threshold approaching 1.5 could Glen Andrew
be selected. In essence, a larger threshold gives more Director - Sales
weight to the network that indicates a no high imped- (205)613-7269 USA
[email protected]
ance fault situation. Table 1 summarizes the perfor-
mance as this threshold is varied.

©2008 www.OsmoseUtilities.com
Industry Journal 12