Page 11 - CE_Industral_Journal_Jan2008

A Novel Approach for Ground Fault Detection

Ground Fault Detection using Neural Neural Networks have been successfully used in many
Network applications to solve complex classification prob-
Figure 5 is a flowchart showing neural network based lems due to their ability to create non-linear decision
high impedance ground fault detection system. boundaries. The most common and flexible neural
Acquired data is filtered using a bandpass filter. The network architecture is the multi-layer backpropaga-
samples are transformed using a fast Fourier transform tion which is constructed from a series of neurons.
(FFT) which is used only in the second neural network The first neural network investigated used 1000 input
described below, and then mapped into the high nodes to the network. No attempt was made to syn-
impedance fault detection plane using the neural chronize the zero crossing of the monitored current
network algorithm and compared to a threshold to de- to the first input node of the neural network with the
termine if a high impedance ground fault has occurred. hopes of reducing implementation complexity. The
best results occurred when using 200 nodes in the
hidden layer. The network was trained with 600 input/
target cases (300 high impedance fault data and 300
Data Acquisition load data) and had a sum-squared error of 1.4 after
completion of learning (1 missed detection and 0 false
alarms). Generalization performance was determined
by testing the network on 3600 new 3 cycle windows
Data Filtering (1800 high impedance faults and 1800 loads). Consid-
ering all network output greater than 0.5 to indicate
the presence of a high impedance ground fault, the
network achieved a detection rate of 70.83% with a
FFT Transformation 22.06% false alarm rate.

Neutral Network
Algorithm

Comparison with
Threshold

Detection Decision

Figure 5. Neural network based high impedance
ground fault detection system.

Industry Journal 10