Instrumentation & Measurement Magazine 24-2 - 105

Table 3 - Comparison between acoustic and RF methods
Acoustic detection

RF detection

Sensor details

MK-720 with 40 kHz as the center frequency
with a frequency range of 16 kHz - 80 kHz

Antenna with a pass band between 1-2 GHz and
a corresponding gain in the range 14.5-18 dB

Selected features

60, 120 and 180 Hz of the captured envelope of
the acoustic signal

Entropy, energy, kurtosis and skewness for the
wavelet packet decomposition

Detected problems

Corona, crack in the insulator and surface
discharge (3 classes)

Corona, crack and hole in the insulator
(3 classes)

Classification accuracy

88-90%

92-98%

Fig. 4. Overall process of machine learning application using both the RF antenna and acoustic sensor from the data collection to the classification.

digital image processing and intelligent techniques, to remove
the subjectivity introduced into the classification by humans,
and to reduce the risk involved in inspection of insulators next
to elevated high voltage lines. Most of the existing methods
used the area ratio of water droplets, the shape factor of water
droplets, or a combination of both with sophisticated mathematical manipulations to obtain accurate results. For example,
the triangular module operator, along with the triangular fusion method, is used to combine the two image features and
consequently get the hydrophobicity class [8]. Another example is using the fusion decision algorithm to combine the
results of three different classifiers (using D-S algorithm) [9].
Other hybrid systems used the same image features with other
non-image distinguishing features to gain better accuracy in
classification results. An example of this is the addition of dielectric measurement as a third feature, which also varies with
hydrophobicity class [10]. The previous studies have concentrated on developing the classification algorithms, but they
lack the development of the hardware system that can be used
in the field to automate the whole process.
April 2021

Detection of Pollution in Outdoor Insulators
The pollution severity is usually quantified using both ESDD
and non-soluble deposit density (NSDD). Estimating the
pollution levels can help utilities arrange washing and maintenance schedules. Measurement of the ESDD and NSDD is
intrusive as it requires the removal of the insulator from the
network, and hence, power shutdown is inevitable. Therefore, there is a need to measure or predict the ESDD using
non-intrusive techniques which requires the use of ML. Most
of the reported studies were performed on ceramic insulators
[11]-[15]. Both image related features and ANN were used to
investigate the ability of the proposed algorithm to establish
a correlation image features and different ESDD levels [10].
Based on the contamination severity, the ESDD was divided
into four classes (low, medium, heavy and very heavy), and
two types of image features (histogram and singular value
decomposition (SVD) based features) were extracted as input feature vectors. A relatively high recognition rate of 86%
was achieved using the proposed algorithm. However, using image features to estimate ESDD can only be feasible if the

IEEE Instrumentation & Measurement Magazine 105

Instrumentation & Measurement Magazine 24-2

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