Instrumentation & Measurement Magazine 24-2 - 106

Fig. 5. Overall process of machine learning application using images for NSDD class prediction.

ESDD and the NSDD are highly correlated which may not be
always the case. In other words, image-based features actually
are effective to directly classify NSDD and indirectly classify
ESDD when both of these pollution metrics are themselves
correlated.
Another indirect and more efficient method to predict
the ESDD is by measuring the surface discharges which
correlate with the severity of ESDD. ANN was utilized to
predict the ESDD level using features extracted from ultrasound emission signals [12]. However, this approach suffers
from the fact that the model was trained on only one distance between the sensor and the insulator. It is well known
that the measured ultrasound signal is distant dependent,
so conducting all measurements at one distance is a limitation to generalize the proposed technique. The leakage
current (LC) peak and third harmonic components, along
with different environmental conditions, have been used to
predict the ESDD level with a high recognition rate of 95.7%
[13]. However, one of the shortcomings of this attempt is
that the proposed system was applied only on one insulator
sample, and hence, generalizing the results is not feasible. The ESDD was predicted using the measured surface
conductivity for glass samples where a linear relationship
was established between the two parameters [14]. Such approach requires that the conductivity measurements be
conducted on the sample while the insulator is not energized, and hence, the proposed method is as intrusive as
measuring its ESDD level. The peak, average and variance
of LC have been used to predict the ESDD level using ANN
for two different strings of ceramic insulators. The contamination was applied artificially using a clean fog test [15].
The study results were mixed as in some cases the recognition rate was more than 90% and in other cases was less than
70%. It is worth mentioning that all approaches that utilize
the surface LC to predict ESDD require that the pollutant is
wet, and hence, these techniques cannot be applied in the
field during dry weather conditions.
106

The previously discussed research concentrated on predicting the ESDD level on both glass and porcelain insulators.
However, few studies were conducted to predict the ESDD
level for non-ceramic insulators [16]. In one study, several
features were extracted from the measured LC during a salt
fog test to predict both the value and class of ESDD level. To
reduce the length of the input feature vector, two feature reduction techniques were used for the prediction: the stepwise
regression and PCA analysis, and to investigate the efficiency
of different classifiers, three different classifiers were utilized,
namely, KNN, polynomial and Neuro-fuzzy classifiers. It has
been found that both the feature reduction technique and
classifier type is crucial to increase the recognition rate. The
proposed technique can be utilized online to predict the severity of the pollution level using the measured LC components
as an input feature.
While predicting the ESDD can be done mainly by measuring other parameters like LC, NSDD could only be predicted
by estimating its value/class directly. Hence, classification of
NSDD requires the analysis of images taken for polluted insulators. Recently, an attempt was conducted to estimate the
level of NSDD using extracted features from the captured images. Only two features (the global mean and area of the first
80 tonal values) extracted from the gray scale of the captured
images were used as the input feature vectors. It is worth
mentioning that only two features were able to generate four
separable classes of NSDD that resulted in classification accuracy more than 90% [17]. A summary of the ML process is
shown in Fig. 5.

What is Next?
It is evident from the previous discussions that applying ML in
insulation diagnostics has a great potential. To further increase
the robustness of ML to detect different problems, the following points need to be further investigated:
◗◗ Increase the application of ML under field conditions:
Most of the reported research was conducted under

IEEE Instrumentation & Measurement Magazine

April 2021

Instrumentation & Measurement Magazine 24-2

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-2