design of anfis for hydrophobicity classification of
TRANSCRIPT
energies
Article
Design of ANFIS for Hydrophobicity Classification ofPolymeric Insulators with Two-Stage FeatureReduction Technique and Its Field Deployment
Rajamohan Jayabal * , K. Vijayarekha and S. Rakesh Kumar
Department of Electrical and Electronics Engineering, School of Electrical and Electronics Engineering, SASTRADeemed University, Thanjavur 613401, India; [email protected] (K.V.); [email protected] (S.R.K.)* Correspondence: [email protected]; Tel.: +91-9952-6870-56
Received: 13 November 2018; Accepted: 28 November 2018; Published: 4 December 2018 �����������������
Featured Application: This work describes the design of an intelligent tool for hydrophobicityclassification of polymeric insulators used in electrical transmission lines. It eliminates themanual inspection of insulators, which involves a significant amount of labor work load.This automated tool can be integrated to an unmanned aerial vehicle to provide autonomousinspection of insulators.
Abstract: Hydrophobicity of polymeric insulator plays a vital role in determining the insulationquality in outdoor overhead electrical transmission and distribution lines. Loss of hydrophobicityincreases the leakage current and leads to flashover. Monitoring hydrophobicity becomes afundamental requirement to ensure continuity of power line operations. Hydrophobicity of polymericinsulator is classified according to STRI (Swedish Transmission Research Institute) guidelines. Thispaper proposes an intelligent ANFIS (Adaptive Neuro-Fuzzy Inference System) based classifier todetermine the hydrophobicity quality using the digital image of the insulator. Ten statistical featuresare extracted from the digital images. Two stages of feature reduction are employed to reduce thenumber of features. Pre-design stage uses PCA (Principal Component Analysis) and reduces thenumber of features to six from ten and the post-design stage analyzes the accumulation effect toreduce the number of features to four. Various ANFIS classifiers are trained using these reducedfeatures extracted from the image. The performance of these ANFIS classifiers is evaluated in bothfield and laboratory specimens. Results indicate classification accuracy of 96.4% and 93.3% duringthe training and testing phase when triangular membership function with linear output function isemployed in ANFIS. A GUI (Graphical User Interface) has also been designed to facilitate the use ofthe proposed system by field operators.
Keywords: polymeric insulators; hydrophobicity; image processing; PCA; ANFIS; GUI
1. Introduction
Electrical insulator plays a vital role in the electrical transmission and distribution system [1].Outdoor insulators are monitored and evaluated to know the aging performance of insulators therebyassuring continuous delivery of power. High voltage polymeric insulators are preferred nowadaysdue to its high hydrophobic surface, faster hydrophobicity recovery, and lower weight over the glassand ceramic insulators [2]. Hydrophobicity reduces with aging, temperature, and pollution depositson the surface [3–5], which leads to a flow of more leakage current, arcing on the surface, and flashoverphenomena [6–8]. Therefore, periodic hydrophobicity measurement is essential to protect and safelyoperate the transmission lines.
Energies 2018, 11, 3391; doi:10.3390/en11123391 www.mdpi.com/journal/energies
Energies 2018, 11, 3391 2 of 16
Various methods are available in the literature to evaluate the hydrophobic condition ofinsulators [9,10]. Contact angle, surface tension, and spray method are the three widely used methodsaccording to the standard IEC 62073 [11]. In the contact angle method, the contact angle betweenthe water droplet and the insulator surface is measured, which is a direct marker of hydrophobicity.Surface tension of the water droplets over the insulator is measured to evaluate its hydrophobicitycondition. These two methods demand sophisticated apparatus and laboratory conditions making itnot suitable for field studies.
In the spray method, water is sprayed on the insulator surface, and the dispersion of waterdroplets is analyzed. It makes this method simple, with low hardware requirements. It is a widelyadopted method for both laboratory and field studies. Dispersion of water droplets is analyzedusing a digital image processing method [12]. It involves pre-processing, feature extraction andclassification. Pre-processing prepares and enhances the digital image to show the water dispersionpattern in a better way. In the literature, histogram equalization and top-hat filters [13,14] are used toprocess the image to compensate for ambient light and to remove noise respectively. Features suchas fractal dimension, shape factor of the droplet, area ratio, haralick texture descriptors and staticcontact angle [15–17] are extracted from the digital image. As these features are an indirect markerof hydrophobicity, the literature shows that researchers have extracted more features from the imageto increase accuracy. It leads to the extraction of dependent and highly correlated features, whichincreases the computational overhead and coveys no new information. Thus, selection of optimal andindependent features is vital and a challenging problem.
Hydrophobicity classification employs both classical methods and learning techniques. Classicalmethods include linear discriminant analysis, support vectors, decision trees and nearest neighborhoodbased classifiers which are designed based on the variations in features [18–21]. Neural networks(NN) and fuzzy clustering are widely adopted intelligent techniques. NN has a learning capabilityto learn the correlation between the features and its corresponding hydrophobicity class. However,the unbounded nature of NN demands tightly bounded features for stable operation. As the imagesare acquired from various locations, variations in the ambient environment are inevitable. This causesfeatures to have a broader range of variations, which causes NN to be unstable. On the other hand,fuzzy systems are robust and produce stable outputs. However, they require manual tuning of rules,which demands expert’s knowledge.
This problem can be overcome by employing ANFIS for hydrophobicity classification. ANFIScombines the learning capability of NN and the efficient knowledge representation potential of a FuzzyInference System (FIS). NN is used to learn the nature of feature variation across the different class ofinsulators. This learning is imposed into the FIS in the form of the rule base, and parameters are tunedin accordance. Thus, the robustness of FIS in combination with the learning nature of neural networksmakes the ANFIS the optimal tool for hydrophobicity classification [22].
To the best of our knowledge, this is a first attempt to use ANFIS to classify hydrophobicity.Hydrophobicity of the insulator is evaluated in terms of its water retention capability. The experimentalsetup to evaluate water retention capability is shown in Figure 1. Initially, the distilled water is sprayedon the surface of insulator uniformly. Images are captured after the spraying is over (first frame).Acquired images are pre-processed to extract the region of interest and to compensate for ambientdisturbances. First and second order statistical features are extracted from the pre-processed image [23].A trained ANFIS classifier is used to classify the hydrophobicity based upon the features. The ANFISrepresents a class of hydrophobicity as numeric. This numeric value is quantized to a particular classof hydrophobicity in post-processing.
Significant contribution of the research article is as follows: (i) Pre-design stage feature reductionusing PCA. (ii) Design and analysis of various ANFIS classifiers. (iii) Post-design stage featurereduction using accumulation effect. (iv) Hydrophobicity classification for laboratory and fieldoperation insulator. (v) Field deployment using GUI to facilitate the operator.
Energies 2018, 11, 3391 3 of 16Energies 2018, 11, x FOR PEER REVIEW 3 of 17
Figure 1. Proposed hydrophobicity classifier.
Significant contribution of the research article is as follows: (i) Pre-design stage feature
reduction using PCA. (ii) Design and analysis of various ANFIS classifiers. (iii) Post-design stage
feature reduction using accumulation effect. (iv) Hydrophobicity classification for laboratory and
field operation insulator. (v) Field deployment using GUI to facilitate the operator.
The research article is organized as follows: Section 2 describes the experimental procedure and
image acquisition techniques, Section 3 describes the design of various ANFIS classifier and feature
reduction techniques, Section 4 describes the performance evaluation and results for laboratory and
field specimens and Section 5 concludes the work with a summary and future directions.
2. Experimental Setup
In this experimental setup, a camera was employed to capture the image of the polymer
insulator. Insulator material considered in this study was grey of type HTVSIR (High-Temperature
Vulcanized Silicone Rubber). Samples were cut off from the new polymer insulator. A solution
prepared with the different concentration of isopropyl alcohol mixed with distilled water [24] was
sprayed on the surface of a new polymer insulator. It closely replicates the aging behavior of the
polymeric insulator deployed in the field. As the polymer ages, its hydrophobic property degrades,
and the same phenomena occur with the fresh polymer insulator when the concentration of
isopropyl alcohol is increased.
Experiments were carried out as per the Swedish Transmission Research Institute (STRI) guide
[25] as follows. In the proposed work, a common spray bottle was chosen to produce a fine mist of
water. A square surface of the insulator with a side 8 cm was chosen, which gives the test area of 64
cm2. About 20 sprays in the span of 20 s were sprayed over the insulator surface. After 5 s, the
images were acquired, and features were extracted for hydrophobicity classification.
The camera was mounted vertically over the insulator at a distance of 25 ± 2 cm and set on
normal mode. This set up was maintained for acquisition of all images in the laboratory and field
test. It enables the capture of images with clear visibility of the insulator surface with solution
droplets with uniform scaling.
Image Acquisition and Pre-Processing
The spraying solutions were prepared with different levels of concentration ranging from 0% to
100% for every 10% step of isopropyl alcohol. This gives about 10 different classes of solution
concentration. A minimum of 30 images from each class of concentration were acquired, and a total
of 681 images were captured. These images were converted to grey scale, which enables the solution
droplets to be more visible over the insulator surface. Then, the images were cropped to select the
region of interest (ROI) and to maintain uniform image size. Figure 2 shows the grey converted and
cropped image for 0% to 100% volume of alcohol with distilled water solution and its corresponding
HC (Hydrophobicity Class). It was observed that the solution droplets are more distinct for the low
level of alcohol concentration and that they get dispersed at a higher concentration. This indicates
that the solution sprayed acts as an excellent marker for hydrophobicity classification.
Figure 1. Proposed hydrophobicity classifier.
The research article is organized as follows: Section 2 describes the experimental procedure andimage acquisition techniques, Section 3 describes the design of various ANFIS classifier and featurereduction techniques, Section 4 describes the performance evaluation and results for laboratory andfield specimens and Section 5 concludes the work with a summary and future directions.
2. Experimental Setup
In this experimental setup, a camera was employed to capture the image of the polymer insulator.Insulator material considered in this study was grey of type HTVSIR (High-Temperature VulcanizedSilicone Rubber). Samples were cut off from the new polymer insulator. A solution prepared with thedifferent concentration of isopropyl alcohol mixed with distilled water [24] was sprayed on the surfaceof a new polymer insulator. It closely replicates the aging behavior of the polymeric insulator deployedin the field. As the polymer ages, its hydrophobic property degrades, and the same phenomena occurwith the fresh polymer insulator when the concentration of isopropyl alcohol is increased.
Experiments were carried out as per the Swedish Transmission Research Institute (STRI) guide [25]as follows. In the proposed work, a common spray bottle was chosen to produce a fine mist of water.A square surface of the insulator with a side 8 cm was chosen, which gives the test area of 64 cm2.About 20 sprays in the span of 20 s were sprayed over the insulator surface. After 5 s, the images wereacquired, and features were extracted for hydrophobicity classification.
The camera was mounted vertically over the insulator at a distance of 25 ± 2 cm and set onnormal mode. This set up was maintained for acquisition of all images in the laboratory and field test.It enables the capture of images with clear visibility of the insulator surface with solution droplets withuniform scaling.
Image Acquisition and Pre-Processing
The spraying solutions were prepared with different levels of concentration ranging from 0%to 100% for every 10% step of isopropyl alcohol. This gives about 10 different classes of solutionconcentration. A minimum of 30 images from each class of concentration were acquired, and a totalof 681 images were captured. These images were converted to grey scale, which enables the solutiondroplets to be more visible over the insulator surface. Then, the images were cropped to select theregion of interest (ROI) and to maintain uniform image size. Figure 2 shows the grey converted andcropped image for 0% to 100% volume of alcohol with distilled water solution and its correspondingHC (Hydrophobicity Class). It was observed that the solution droplets are more distinct for the lowlevel of alcohol concentration and that they get dispersed at a higher concentration. This indicates thatthe solution sprayed acts as an excellent marker for hydrophobicity classification.
Energies 2018, 11, 3391 4 of 16Energies 2018, 11, x FOR PEER REVIEW 4 of 17
Figure 2. Images of the insulator with varying % concentration and hydrophobicity class. (a) 0%
(HC-1); (b) 10% (HC-2); (c) 20% (HC-2); (d) 30% (HC-3); (e) 40% (HC-3); (f) 50% (HC-4); (g) 60%
(HC-4); (h) 70% (HC-5); (i) 80% (HC-5); (j) 90% (HC-6); (k) 100% (HC-7).
An experienced inspector is able to classify the wettability of the insulator by comparing the
reference image with the acquired image as in Figure 2 and description provided in the STRI guide
as described in Table 1. Lack of experienced inspectors and the need for periodic inspection of
insulators demands an automation tool to determine the hydrophobicity class. The uncertain nature
in the dispersion of solution demands intelligent techniques to extract inferences about the insulator
quality. On the other hand, the features extracted from images can provide valuable information and
need machine learning techniques to understand variations in features across different class of
hydrophobicity. This makes ANFIS to be ideal for hydrophobicity classification using image
processing technique.
Table 1. Criteria for evaluation of HC.
HC Description
1 Only discrete droplets are formed. θ = 80° or larger for the majority of droplets
2 Only discrete droplets are formed. 50° < θ < 80° for the majority of droplets.
3 Only discrete droplets are formed. 20° < θ < 50° for the majority of droplets.
Usually, they are no longer circular.
4
Both discrete droplets and wetted traces from the water runnels are observed.
(i.e., = 0°). Completely wetted areas <2 cm2.
Together they cover <90% of the tested area.
5 Some completely wetted areas >2 cm2, which cover <90% of the tested area.
6 Wetted areas cover >90%, i.e., small un-wetted areas (spots/traces) are still observed.
7 Continuous water film over the whole tested area.
3. Design of ANFIS Classifier
The design of the ANFIS classifier involves various processes as shown in Figure 3. Images are
acquired and pre-processed as explained in the previous section. In this section, other design steps
are explained briefly.
Figure 2. Images of the insulator with varying % concentration and hydrophobicity class. (a) 0%(HC-1); (b) 10% (HC-2); (c) 20% (HC-2); (d) 30% (HC-3); (e) 40% (HC-3); (f) 50% (HC-4); (g) 60% (HC-4);(h) 70% (HC-5); (i) 80% (HC-5); (j) 90% (HC-6); (k) 100% (HC-7).
An experienced inspector is able to classify the wettability of the insulator by comparing thereference image with the acquired image as in Figure 2 and description provided in the STRI guideas described in Table 1. Lack of experienced inspectors and the need for periodic inspection ofinsulators demands an automation tool to determine the hydrophobicity class. The uncertain naturein the dispersion of solution demands intelligent techniques to extract inferences about the insulatorquality. On the other hand, the features extracted from images can provide valuable informationand need machine learning techniques to understand variations in features across different classof hydrophobicity. This makes ANFIS to be ideal for hydrophobicity classification using imageprocessing technique.
Table 1. Criteria for evaluation of HC.
HC Description
1 Only discrete droplets are formed. θ = 80◦ or larger for the majority of droplets2 Only discrete droplets are formed. 50◦ < θ < 80◦ for the majority of droplets.
3 Only discrete droplets are formed. 20◦ < θ < 50◦ for the majority of droplets.Usually, they are no longer circular.
4Both discrete droplets and wetted traces from the water runnels are observed.(i.e., = 0◦). Completely wetted areas <2 cm2.Together they cover <90% of the tested area.
5 Some completely wetted areas >2 cm2, which cover <90% of the tested area.6 Wetted areas cover >90%, i.e., small un-wetted areas (spots/traces) are still observed.7 Continuous water film over the whole tested area.
3. Design of ANFIS Classifier
The design of the ANFIS classifier involves various processes as shown in Figure 3. Images areacquired and pre-processed as explained in the previous section. In this section, other design steps areexplained briefly.
Energies 2018, 11, 3391 5 of 16
Energies 2018, 11, x FOR PEER REVIEW 5 of 17
Figure 3. Steps in the design of the ANFIS classifier.
3.1. Extraction of Feature
A total of ten, first and second order statistical features were extracted from the grey scale
image. First order features provide inferences about the characteristics of individual pixels
cumulatively. Average intensity, standard deviation, variance, skewness, and kurtosis are the first
order features considered in this study. Second order features like entropy, homogeneity, contrast,
correlation, and energy were also extracted. These features provide inferences about the variations
of pixels in contrast with its neighborhood, which is vital for determining the distribution of
droplets.
3.2. Pre-Stage Feature Reduction Using PCA
From the features mentioned, best features are selected by principal component analysis (PCA).
The best features are characterized by the quality of inference and the correlation across the other
features. PCA is a widely used technique to reduce the dimensionality of the feature set by
maintaining the correlation structure [26]. This enables the classifier to classify hydrophobicity with
reduced features and minimizing computational overheads.
Eigen analysis of the correlation matrix has been carried out as described in Table 2. The
proportional Eigen value demonstrates the importance of a particular principal component (PC). It
also describes the proportion of the feature data set covered by the corresponding PC. For instance,
the PC1 and PC2 covers up 60.06% and 17.29% of data set, respectively. This makes the first two
principal components (PC1 and PC2) account for 0.7735 (77.35%) of total data set as described by
cumulative Eigen values in Table 2. Therefore, Eigen vectors corresponding to the first two PC’s are
Figure 3. Steps in the design of the ANFIS classifier.
3.1. Extraction of Feature
A total of ten, first and second order statistical features were extracted from the grey scale image.First order features provide inferences about the characteristics of individual pixels cumulatively.Average intensity, standard deviation, variance, skewness, and kurtosis are the first order featuresconsidered in this study. Second order features like entropy, homogeneity, contrast, correlation,and energy were also extracted. These features provide inferences about the variations of pixels incontrast with its neighborhood, which is vital for determining the distribution of droplets.
3.2. Pre-Stage Feature Reduction Using PCA
From the features mentioned, best features are selected by principal component analysis (PCA).The best features are characterized by the quality of inference and the correlation across the otherfeatures. PCA is a widely used technique to reduce the dimensionality of the feature set by maintainingthe correlation structure [26]. This enables the classifier to classify hydrophobicity with reducedfeatures and minimizing computational overheads.
Eigen analysis of the correlation matrix has been carried out as described in Table 2.The proportional Eigen value demonstrates the importance of a particular principal component(PC). It also describes the proportion of the feature data set covered by the corresponding PC. Forinstance, the PC1 and PC2 covers up 60.06% and 17.29% of data set, respectively. This makes the firsttwo principal components (PC1 and PC2) account for 0.7735 (77.35%) of total data set as described
Energies 2018, 11, 3391 6 of 16
by cumulative Eigen values in Table 2. Therefore, Eigen vectors corresponding to the first two PC’sare considered as described in Table 3 and other PC’s are made obsolete because of being inferiorand interdependencies. Best features were selected based on a magnitude threshold as highlighted inTable 3. A scree plot shown in Figure 4 illustrates the distribution of Eigen values for the feature setused in this study. Thus, the use of PCA reduces the features required for hydrophobicity classificationfrom 10 to six features.
Table 2. Eigen analysis of the correlation matrix.
PC Eigen Value Proportion Cumulative
PC1 6.0064 0.6006 0.6006PC2 1.7286 0.1729 0.7735PC3 0.9479 0.0948 0.8683PC4 0.5684 0.0568 0.9251PC5 0.4307 0.0431 0.9682PC6 0.2336 0.0234 0.9915PC7 0.0648 0.0065 0.9981PC8 0.0165 0.0016 0.9997PC9 0.0028 0.0003 0.9999PC10 0.0004 4.39 × 10−5 1.0000
Table 3. Eigen vectors of PC1 and PC2.
Features Feature Order PC1 PC2
Average Intensity 1 0.1384 0.5761Variance - 0.3903 −0.1003Skewness 3 0.1460 0.4520Kurtosis - −0.1930 −0.2235
Standard deviation 5 0.3979 −0.1426Entropy - 0.3732 −0.2268
Homogeneity 4 −0.4027 0.0243Contrast 6 0.3959 −0.0392
Correlation 2 −0.1144 −0.5604Energy - −0.3747 0.1257
Energies 2018, 11, x FOR PEER REVIEW 6 of 17
considered as described in Table 3 and other PC’s are made obsolete because of being inferior and
interdependencies. Best features were selected based on a magnitude threshold as highlighted in
Table 3. A scree plot shown in Figure 4 illustrates the distribution of Eigen values for the feature set
used in this study. Thus, the use of PCA reduces the features required for hydrophobicity
classification from 10 to six features.
Table 2. Eigen analysis of the correlation matrix.
PC Eigen Value Proportion Cumulative
PC1 6.0064 0.6006 0.6006
PC2 1.7286 0.1729 0.7735
PC3 0.9479 0.0948 0.8683
PC4 0.5684 0.0568 0.9251
PC5 0.4307 0.0431 0.9682
PC6 0.2336 0.0234 0.9915
PC7 0.0648 0.0065 0.9981
PC8 0.0165 0.0016 0.9997
PC9 0.0028 0.0003 0.9999
PC10 0.0004 4.39 × 10−5 1.0000
Table 3. Eigen vectors of PC1 and PC2.
Features Feature Order PC1 PC2
Average Intensity 1 0.1384 0.5761
Variance - 0.3903 −0.1003
Skewness 3 0.1460 0.4520
Kurtosis - −0.1930 −0.2235
Standard deviation 5 0.3979 −0.1426
Entropy - 0.3732 −0.2268
Homogeneity 4 −0.4027 0.0243
Contrast 6 0.3959 −0.0392
Correlation 2 −0.1144 −0.5604
Energy - −0.3747 0.1257
Figure 4. Scree plot of Eigen values.
3.3. Training of the ANFIS Classifier
A total of 681 images under the various classes of hydrophobicity were used in this study. Six
best features were extracted for these images, and the corresponding hydrophobicity class was
Figure 4. Scree plot of Eigen values.
3.3. Training of the ANFIS Classifier
A total of 681 images under the various classes of hydrophobicity were used in this study. Six bestfeatures were extracted for these images, and the corresponding hydrophobicity class was assigned
Energies 2018, 11, 3391 7 of 16
to them. A total of 591 images were used to generate features for training the ANFIS classifier,and 90 images were considered for the testing phase. ANFIS is a hybridization of the neural and fuzzysystem. Takagi-Sugeno type fuzzy inference system (FIS) is used in ANFIS [27,28]. ANFIS uses theartificial neural network (ANN) to learn knowledge from the feature set, and the learning is imposedinto the FIS. It is composed of five layers, as shown in Figure 5 and the operation of each layer isdescribed in Table 4. Out of five layers, two layers namely ‘Input Layer’ and ‘Defuzzification Layer’were considered as an adaptive layer. The parameters of these adaptive layers are tuned using theANN to represent the knowledge learned from the training dataset [29].
Energies 2018, 11, x FOR PEER REVIEW 7 of 17
assigned to them. A total of 591 images were used to generate features for training the ANFIS
classifier, and 90 images were considered for the testing phase. ANFIS is a hybridization of the
neural and fuzzy system. Takagi-Sugeno type fuzzy inference system (FIS) is used in ANFIS [27,28].
ANFIS uses the artificial neural network (ANN) to learn knowledge from the feature set, and the
learning is imposed into the FIS. It is composed of five layers, as shown in Figure 5 and the operation
of each layer is described in Table 4. Out of five layers, two layers namely ‘Input Layer’ and
‘Defuzzification Layer’ were considered as an adaptive layer. The parameters of these adaptive
layers are tuned using the ANN to represent the knowledge learned from the training dataset [29].
Figure 5. Architecture of ANFIS.
Table 4. Description of ANFIS.
ANN
Layers Layer Name FIS Functions Description
Layer-1 Input Layer Performs Fuzzification
Adaptive Nodes—Adapt the Premise
Parameters of the trapezoidal
membership function (a, b, c, d) and
triangular membership function (a, b, c)
Layer-2 Membership
layer
Computes the firing
strength of rules Non-Adaptive Nodes
Layer-3 Rule Layer
Computes the
normalized firing
strength for each
rule (node)
Non-Adaptive Nodes
Layer-4 Defuzzification
Layer
Provides output values
resulting from the
inference of rules
Adaptive Nodes—Adapt the
Consequent Parameters (p, q, r) of the
rules ith Rule of Rule Base:
1 21 2
1 2
mf
n n
mf
i i i i
if X isX AND X isX
then f p X q X r
Layer-5 Output Layer
Sums up all the output
values to create the
crisp output
Non-Adaptive Nodes
Figure 5. Architecture of ANFIS.
Table 4. Description of ANFIS.
ANN Layers Layer Name FIS Functions Description
Layer-1 Input Layer Performs Fuzzification
Adaptive Nodes—Adapt the PremiseParameters of the trapezoidal
membership function (a, b, c, d) andtriangular membership function (a, b, c)
Layer-2 Membershiplayer
Computes the firingstrength of rules Non-Adaptive Nodes
Layer-3 Rule Layer
Computes thenormalized firing
strength for each rule(node)
Non-Adaptive Nodes
Layer-4 DefuzzificationLayer
Provides output valuesresulting from theinference of rules
Adaptive Nodes—Adapt theConsequent Parameters (p, q, r) of the
rules ith Rule of Rule Base:i f X1 is Xn
1m f and X2 is Xn2m f
then fi = piX1 + qiX2 + ri
Layer-5 Output LayerSums up all the output
values to create thecrisp output
Non-Adaptive Nodes
The input layer corresponds to the fuzzification operation of the inputs. It is responsible for theassignment of degree of membership to the crisp value of features, which is a critical parameter inclassification problems. From the literature, it is observed that triangular and trapezoidal membership
Energies 2018, 11, 3391 8 of 16
functions (MF) are widely deployed for classification problems [30]. Triangular membership functionhas a peak membership for a single value of the variable, and the degree of membership degrades asthe value of variable deviates. The degradation in membership function can be controlled by threetunable parameters (a, b, c), which determine the shape of the membership function as shown inFigure 6a. These parameters are tuned by the learning algorithm based on the variations of the features.
Energies 2018, 11, x FOR PEER REVIEW 8 of 17
The input layer corresponds to the fuzzification operation of the inputs. It is responsible for the
assignment of degree of membership to the crisp value of features, which is a critical parameter in
classification problems. From the literature, it is observed that triangular and trapezoidal
membership functions (MF) are widely deployed for classification problems [30]. Triangular
membership function has a peak membership for a single value of the variable, and the degree of
membership degrades as the value of variable deviates. The degradation in membership function
can be controlled by three tunable parameters (a, b, c), which determine the shape of the
membership function as shown in Figure 6a. These parameters are tuned by the learning algorithm
based on the variations of the features.
Figure 6. Input membership functions (MF) used for ANFIS design. (a) triangular MF; (b)
trapezoidal MF.
Like triangular membership functions, trapezoidal MF also has a linearly degrading
membership, and they also have a stable membership degree to a certain range of features. The
stable range is determined by the parameters (b, c) and the degradation in membership is
determined by (a, d) parameters as in Figure 6b. This makes the triangular MF an ideal tool for
classifier with the quantized class of hydrophobicity. Thus, in the proposed work, triangular and
trapezoidal membership functions are considered for designing ANFIS classifier.
The defuzzification layer is responsible for the integration of all the fired rules and conversion
of fuzzy value to a crisp value corresponding to hydrophobicity class. Two types of defuzzification
functions namely constant and linear functions were investigated in this work as described in Table
5. The constant function produces a constant value when the corresponding rule is fired whereas the
linear function evaluates the linear combination of all the input membership values to produce a
crisp output. The design procedure involves preparation and loading of a feature set, selection of
ANFIS structure (type and number of MF’s), selection of training algorithms, and execution of the
training algorithm with stopping criteria as illustrated in Figure 7.
Table 5. Various types of ANFIS classifier.
ANFIS Type Input MF Type Type of Output Function
Classifier-1 Triangular Constant
Classifier-2 Triangular Linear
Classifier-3 Trapezoidal Constant
Classifier-4 Trapezoidal Linear
Figure 6. Input membership functions (MF) used for ANFIS design. (a) triangular MF;(b) trapezoidal MF.
Like triangular membership functions, trapezoidal MF also has a linearly degrading membership,and they also have a stable membership degree to a certain range of features. The stable range isdetermined by the parameters (b, c) and the degradation in membership is determined by (a, d)parameters as in Figure 6b. This makes the triangular MF an ideal tool for classifier with the quantizedclass of hydrophobicity. Thus, in the proposed work, triangular and trapezoidal membership functionsare considered for designing ANFIS classifier.
The defuzzification layer is responsible for the integration of all the fired rules and conversionof fuzzy value to a crisp value corresponding to hydrophobicity class. Two types of defuzzificationfunctions namely constant and linear functions were investigated in this work as described in Table 5.The constant function produces a constant value when the corresponding rule is fired whereas thelinear function evaluates the linear combination of all the input membership values to produce a crispoutput. The design procedure involves preparation and loading of a feature set, selection of ANFISstructure (type and number of MF’s), selection of training algorithms, and execution of the trainingalgorithm with stopping criteria as illustrated in Figure 7.
Table 5. Various types of ANFIS classifier.
ANFIS Type Input MF Type Type of Output Function
Classifier-1 Triangular ConstantClassifier-2 Triangular LinearClassifier-3 Trapezoidal ConstantClassifier-4 Trapezoidal Linear
The hybrid learning algorithm [31] was employed for training the ANFIS classifier. It is acombination of back propagation algorithm (BPA) and the least-squares method. The least-squaresmethod uses forward propagation error to estimate initial parameters of layer-1 and 4 (as in Table 4).BPA back-propagates the error to optimize the parameters, which minimize the mean square error(MSE) between the actual and predicted class. Thus, the hybrid learning algorithm can train in a lessernumber of iterations (epochs) as compared to BPA alone as shown in Figure 8.
Energies 2018, 11, 3391 9 of 16
Energies 2018, 11, x FOR PEER REVIEW 9 of 17
(a)
(b)
Figure 7. ANFIS Design Procedure. (a) Loading of training data set; (b) selection of ANFIS structure.
The hybrid learning algorithm [31] was employed for training the ANFIS classifier. It is a
combination of back propagation algorithm (BPA) and the least-squares method. The least-squares
method uses forward propagation error to estimate initial parameters of layer-1 and 4 (as in Table 4).
BPA back-propagates the error to optimize the parameters, which minimize the mean square error
(MSE) between the actual and predicted class. Thus, the hybrid learning algorithm can train in a
lesser number of iterations (epochs) as compared to BPA alone as shown in Figure 8.
(a)
(b)
Figure 8. Training of the ANFIS classifier. (a) hybrid learning algorithm; (b) back propagation
algorithm.
Figure 7. ANFIS Design Procedure. (a) Loading of training data set; (b) selection of ANFIS structure.
Energies 2018, 11, x FOR PEER REVIEW 9 of 17
(a)
(b)
Figure 7. ANFIS Design Procedure. (a) Loading of training data set; (b) selection of ANFIS structure.
The hybrid learning algorithm [31] was employed for training the ANFIS classifier. It is a
combination of back propagation algorithm (BPA) and the least-squares method. The least-squares
method uses forward propagation error to estimate initial parameters of layer-1 and 4 (as in Table 4).
BPA back-propagates the error to optimize the parameters, which minimize the mean square error
(MSE) between the actual and predicted class. Thus, the hybrid learning algorithm can train in a
lesser number of iterations (epochs) as compared to BPA alone as shown in Figure 8.
(a)
(b)
Figure 8. Training of the ANFIS classifier. (a) hybrid learning algorithm; (b) back propagation
algorithm. Figure 8. Training of the ANFIS classifier. (a) hybrid learning algorithm; (b) backpropagation algorithm.
It was observed that an initial guess of parameters brings the MSE to 0.72 in hybrid learningwhereas 4.5 MSE is observed in BPA. For 40 epochs, the hybrid learning algorithm tends to convergetowards optimal parameters reducing the MSE to 0.28. BPA shows a minimal descending MSE from4.5 to 4.17 for the same 40 epochs.
Energies 2018, 11, 3391 10 of 16
3.4. Testing of ANFIS Classifier
The performance of the proposed hydrophobic classifier was evaluated using the training datasetand testing dataset as illustrated in Figure 9.
Energies 2018, 11, x FOR PEER REVIEW 10 of 17
It was observed that an initial guess of parameters brings the MSE to 0.72 in hybrid learning
whereas 4.5 MSE is observed in BPA. For 40 epochs, the hybrid learning algorithm tends to converge
towards optimal parameters reducing the MSE to 0.28. BPA shows a minimal descending MSE from
4.5 to 4.17 for the same 40 epochs.
3.4. Testing of ANFIS Classifier
The performance of the proposed hydrophobic classifier was evaluated using the training
dataset and testing dataset as illustrated in Figure 9.
(a)
(b)
Figure 9. Testing of the ANFIS classifier. (a) training data set; (b) testing data set.
The outputs of the classifier were designated to be integers ranging from 1 to 7 indicating the
hydrophobicity class. Though the outputs were quantized, the ANFIS classifier produces decimal
points as output (as in Figure 9), and it has to be rounded off to the nearest class of hydrophobicity. It
was observed that the classifier is accurate at a lower class of hydrophobicity and a significant
perturbation is observed across class 5 and 6 for the training dataset. Similar perturbation was also
observed in a testing dataset with minimal loss of accuracy at lower classes.
3.5. Post-Stage Feature Reduction Using the Accumulation Effect
The accumulation effect on classification accuracy is evaluated to investigate further feature
reduction. Classification accuracy was evaluated using a confusion matrix. The classification
accuracy was calculated from the ratio of a number of correct prediction (NC), which is expressed in
Equation (1) to the total number of samples (NS). It is expressed in terms of percentage from the
confusion matrix as in Equation (2).
Figure 9. Testing of the ANFIS classifier. (a) training data set; (b) testing data set.
The outputs of the classifier were designated to be integers ranging from 1 to 7 indicating thehydrophobicity class. Though the outputs were quantized, the ANFIS classifier produces decimalpoints as output (as in Figure 9), and it has to be rounded off to the nearest class of hydrophobicity.It was observed that the classifier is accurate at a lower class of hydrophobicity and a significantperturbation is observed across class 5 and 6 for the training dataset. Similar perturbation was alsoobserved in a testing dataset with minimal loss of accuracy at lower classes.
3.5. Post-Stage Feature Reduction Using the Accumulation Effect
The accumulation effect on classification accuracy is evaluated to investigate further featurereduction. Classification accuracy was evaluated using a confusion matrix. The classification accuracywas calculated from the ratio of a number of correct prediction (NC), which is expressed in Equation (1)to the total number of samples (NS). It is expressed in terms of percentage from the confusion matrixas in Equation (2).
NC = {|PCi|PCi = ACi, 1 < i < NS}, (1)
%Accuracy =NCNS× 100%. (2)
Energies 2018, 11, 3391 11 of 16
ANFIS classifiers are designed by introducing the features in the order as illustrated in Table 3.Initially, the most robust feature (Average Intensity with PC of 0.5761) alone is used as input forhydrophobicity classifier and its accuracy was evaluated using a confusion matrix. Subsequently,the strong feature was added cumulatively to design classifiers. This results in six different inputcombination with increasing accumulated features (nf). This type of feature accumulation wasimplemented on all the four types of classifiers as described in Table 5. It leads to a total of 24 classifiers(6 × 4) whose classification accuracies are evaluated for training and testing feature sets as illustratedin Figure 10a,b respectively.
Energies 2018, 11, x FOR PEER REVIEW 11 of 17
,1C i i i S
N PC PC AC i N , (1)
% 100%C
S
NAccuracy X
N. (2)
ANFIS classifiers are designed by introducing the features in the order as illustrated in Table 3.
Initially, the most robust feature (Average Intensity with PC of 0.5761) alone is used as input for
hydrophobicity classifier and its accuracy was evaluated using a confusion matrix. Subsequently,
the strong feature was added cumulatively to design classifiers. This results in six different input
combination with increasing accumulated features (nf). This type of feature accumulation was
implemented on all the four types of classifiers as described in Table 5. It leads to a total of 24
classifiers (6 × 4) whose classification accuracies are evaluated for training and testing feature sets as
illustrated in Figure 10a,b respectively.
(a)
(b)
Figure 10. Classification accuracy for ANFIS classifier. (a) training data set; (b) testing data set.
It was observed that improvement in classification accuracy becomes insignificantly small with
the addition of 5th and 6th features (standard deviation and contrast). The classification accuracy
saturates beyond using four features. This implies a possibility of a reduction in features to four
without compromising classification accuracy. This reduction in features reduces the computational
Figure 10. Classification accuracy for ANFIS classifier. (a) training data set; (b) testing data set.
It was observed that improvement in classification accuracy becomes insignificantly small withthe addition of 5th and 6th features (standard deviation and contrast). The classification accuracysaturates beyond using four features. This implies a possibility of a reduction in features to fourwithout compromising classification accuracy. This reduction in features reduces the computationalrequirements and processing time. It also facilitates removal of ambiguous features and selection offeatures in coherence to improve the classification accuracy.
Energies 2018, 11, 3391 12 of 16
4. Results and Discussion
The proposed ANFIS classifier was used to analyze both laboratory and field specimens to identifyits hydrophobicity class. The performance of the classifier was analyzed using standard metrics suchas Confusion Matrix, Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), RootMean Square (RMS) value and R2 (Regression) [32–34]. These error metrics were calculated between thepredicted class (PCi) of the ANFIS classifier and the actual class (ACi) designated by the standards/expertinspector. The description and formulae for calculating these measures are given in Table 6.
Table 6. Performance metrics.
Metrics Description Formula
Mean absolute error Average of absolute errors MAE = 1n
n∑
i=1|PCi − ACi|
Mean absolute percentage error Measure of prediction accuracy MAPE = 1n
n∑
i=1
|PCi−ACi |ACi
Root mean square value Square root of the mean square RMSE =
√1n
n∑
i=1(PCi − ACi)
2
R2Proportional measure of variance of
one variable that can is predicted fromthe other variable
R =
√√√√√1−n∑
i=1(PCi−ACi)
2
n∑
i=1AC2
i
4.1. Laboratory Test
In the laboratory, a total of 591 images were used as training data, and 90 images were used astesting data. Classification error was analyzed using a confusion matrix for all the types of input MF’s(triangular and trapezoidal) and output functions (constant and linear) considered in this study. Thedistribution of classes in the confusion matrix for training and testing data of this ANFIS classifier isalso shown in Table 7. Table 8 illustrates the classification accuracy for all the ANFIS designs, and it isobserved that ANFIS with triangular MF and linear function (Classifier-2) exhibits a higher degree ofaccuracy. Further, the performance of the designed ANFIS classifier was also evaluated in terms ofstandard metrics as illustrated in Table 8. The obtained results illustrate a minimal error in classificationfor Classifier-2. The regression (R) is also approaching one indicating a minimal deviation between thepredicted and actual class.
Table 7. Confusion matrix for training and testing data of ANFIS with triangular MF and alinear function.
AC/PC HC-1 HC-2 HC-3 HC-4 HC-5 HC-6 HC-7
HC-1 34 (6) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)HC-2 0 (0) 86 (14) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0)HC-3 0 (0) 0 (0) 104 (16) 0 (1) 0 (0) 0 (0) 0 (0)HC-4 0 (0) 0 (0) 0 (0) 104 (15) 4 (0) 0 (0) 0 (0)HC-5 0 (0) 0 (0) 0 (0) 0 (0) 131 (19) 1 (1) 1 (1)HC-6 0 (0) 0 (0) 0 (0) 0 (0) 5 (1) 51 (7) 10 (2)HC-7 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 0 (0) 60 (7)
()—indicates testing data set; AC—Actual Class; PC—Predicted Class.
Table 8. Performance measures of the ANFIS classifier.
Measure MAE MAPE RMSE R2 %Accuracy
ANFIS Type Training Testing Training Testing Training Testing Training Testing Training Testing
Classifier-1 0.127 0.174 0.038 0.058 0.189 0.266 0.994 0.988 84.8 81.1Classifier-2 0.014 0.094 0.050 0.029 0.076 0.139 0.999 0.997 96.4 93.3Classifier-3 0.166 0.230 0.040 0.057 0.277 0.346 0.987 0.981 88.7 85.6Classifier-4 0.057 0.160 0.014 0.040 0.107 0.323 0.998 0.983 94.6 86.7
Energies 2018, 11, 3391 13 of 16
4.2. Field Test
The performance of the proposed classifier was also evaluated in insulators installed in the livetransmission lines. The test is carried out in ‘Orathanadu Taluk’ of Thanjavur district, Tamil Nadu,India. Three different sites in the same region with outdoor polymeric insulators operating for one year(Site-1) and two years (Site-2 and 3) were considered for this study. Each pole in the test sites has threepolymeric insulators installed, and one of the three insulators were taken for hydrophobicity analysis.
Field tests were carried out during the scheduled maintenance shut down. Two field operatorswere made to climb up the poles near the insulator. One operator was assigned to spraying the distilledwater and other to capture the image using a digital camera as illustrated in Figure 11. Distilledwater was sprayed on the surface of insulator using a sprayer and recorded by the digital camera.Images were captured after the spraying process under three trials. For analysis, five images werecaptured for each trial providing 15 images for each test site. In these frames, only the top layer ofinsulator was considered for hydrophobicity analysis. As these images were acquired in an outdoorenvironment, they were pre-processed with histogram equalization [20] to reduce the variations inambient light illumination.
Energies 2018, 11, x FOR PEER REVIEW 13 of 17
Table 8. Performance measures of the ANFIS classifier.
Measure MAE MAPE RMSE R2 %Accuracy
ANFIS
Type
Train
ing Testing
Train
ing Testing
Train
ing Testing
Train
ing Testing
Train
ing Testing
Classifie
r-1 0.127 0.174 0.038 0.058 0.189 0.266 0.994 0.988 84.8 81.1
Classifie
r-2 0.014 0.094 0.050 0.029 0.076 0.139 0.999 0.997 96.4 93.3
Classifie
r-3 0.166 0.230 0.040 0.057 0.277 0.346 0.987 0.981 88.7 85.6
Classifie
r-4 0.057 0.160 0.014 0.040 0.107 0.323 0.998 0.983 94.6 86.7
4.2. Field Test
The performance of the proposed classifier was also evaluated in insulators installed in the live
transmission lines. The test is carried out in ‘Orathanadu Taluk’ of Thanjavur district, Tamil Nadu,
India. Three different sites in the same region with outdoor polymeric insulators operating for one
year (Site-1) and two years (Site-2 and 3) were considered for this study. Each pole in the test sites
has three polymeric insulators installed, and one of the three insulators were taken for
hydrophobicity analysis.
Field tests were carried out during the scheduled maintenance shut down. Two field operators
were made to climb up the poles near the insulator. One operator was assigned to spraying the
distilled water and other to capture the image using a digital camera as illustrated in Figure 11.
Distilled water was sprayed on the surface of insulator using a sprayer and recorded by the digital
camera. Images were captured after the spraying process under three trials. For analysis, five images
were captured for each trial providing 15 images for each test site. In these frames, only the top layer
of insulator was considered for hydrophobicity analysis. As these images were acquired in an
outdoor environment, they were pre-processed with histogram equalization [20] to reduce the
variations in ambient light illumination.
Figure 11. Field test carried out in the test site.
Classifier-2, which has higher classification accuracy, was used to classify the hydrophobicity
class of these field insulators. After the spraying was done by Operator-1, images of the insulator
surface were acquired by Operator-2. These images were given to an experienced inspector and the
actual class of hydrophobicity was determined by visual inspection. The field insulators were
categorized with hydrophobicity Class 2, 3 and 3 for the samples collected at Sites-1, 2 and 3
respectively. The predicted hydrophobicity classes of these insulators are illustrated in Figure 12. It
was observed that the proposed system can classify the field samples accurately with minimal
variance and reliable accuracy.
Operator2–Image acquisition
Operator 1–Spraying
Figure 11. Field test carried out in the test site.
Classifier-2, which has higher classification accuracy, was used to classify the hydrophobicity classof these field insulators. After the spraying was done by Operator-1, images of the insulator surfacewere acquired by Operator-2. These images were given to an experienced inspector and the actual classof hydrophobicity was determined by visual inspection. The field insulators were categorized withhydrophobicity Class 2, 3 and 3 for the samples collected at Sites-1, 2 and 3 respectively. The predictedhydrophobicity classes of these insulators are illustrated in Figure 12. It was observed that the proposedsystem can classify the field samples accurately with minimal variance and reliable accuracy.Energies 2018, 11, x FOR PEER REVIEW 14 of 17
Figure 12. Hydrophobicity classification in the field test.
4.3. Design of GUI for Field Deployment
A GUI was designed using MATLAB Graphical User Interface Development Environment
(GUIDE) to facilitate the use of the proposed classifier for field testing of insulators as illustrated in
Figure 13. It provides easy access to the functionalities used in the proposed system to the operator.
The image of the insulator was acquired either from the laboratory test condition or field test. The
image can be loaded by the ‘Load Button’ available in the GUI. This ‘Load Image’ button opens a
browser window to select the corresponding image file. Once the image file is successfully loaded,
the image is displayed in the ‘Original Image’ section. It also provides the operator with a facility to
crop the region of interest (ROI) which can be used for analysis of hydrophobicity.
Figure 13. GUI for hydrophobicity classification.
The operator can confirm the ROI by using the ‘Crop Image’ button. The cropped image is
displayed in the ‘Cropped Image’ section for confirmation with the operator. The operator can also
crop the image again, in case of improper selection of ROI. Once the cropped image is finalized, the
class of hydrophobicity is determined using the ‘Classify’ button. The ‘Classify’ button is used to
activate the feature extraction process and evaluation of ANFIS classifier using these extracted
features. The ‘output’ section provides the class of hydrophobicity of the insulator and also remarks
about the condition of the insulator. Thus, the GUI facilities provide easy access to the operator
without any need to learn programme using the proposed ANFIS classifier.
Figure 12. Hydrophobicity classification in the field test.
Energies 2018, 11, 3391 14 of 16
4.3. Design of GUI for Field Deployment
A GUI was designed using MATLAB Graphical User Interface Development Environment(GUIDE) to facilitate the use of the proposed classifier for field testing of insulators as illustratedin Figure 13. It provides easy access to the functionalities used in the proposed system to the operator.The image of the insulator was acquired either from the laboratory test condition or field test. The imagecan be loaded by the ‘Load Button’ available in the GUI. This ‘Load Image’ button opens a browserwindow to select the corresponding image file. Once the image file is successfully loaded, the imageis displayed in the ‘Original Image’ section. It also provides the operator with a facility to crop theregion of interest (ROI) which can be used for analysis of hydrophobicity.
Energies 2018, 11, x FOR PEER REVIEW 14 of 17
Figure 12. Hydrophobicity classification in the field test.
4.3. Design of GUI for Field Deployment
A GUI was designed using MATLAB Graphical User Interface Development Environment
(GUIDE) to facilitate the use of the proposed classifier for field testing of insulators as illustrated in
Figure 13. It provides easy access to the functionalities used in the proposed system to the operator.
The image of the insulator was acquired either from the laboratory test condition or field test. The
image can be loaded by the ‘Load Button’ available in the GUI. This ‘Load Image’ button opens a
browser window to select the corresponding image file. Once the image file is successfully loaded,
the image is displayed in the ‘Original Image’ section. It also provides the operator with a facility to
crop the region of interest (ROI) which can be used for analysis of hydrophobicity.
Figure 13. GUI for hydrophobicity classification.
The operator can confirm the ROI by using the ‘Crop Image’ button. The cropped image is
displayed in the ‘Cropped Image’ section for confirmation with the operator. The operator can also
crop the image again, in case of improper selection of ROI. Once the cropped image is finalized, the
class of hydrophobicity is determined using the ‘Classify’ button. The ‘Classify’ button is used to
activate the feature extraction process and evaluation of ANFIS classifier using these extracted
features. The ‘output’ section provides the class of hydrophobicity of the insulator and also remarks
about the condition of the insulator. Thus, the GUI facilities provide easy access to the operator
without any need to learn programme using the proposed ANFIS classifier.
Figure 13. GUI for hydrophobicity classification.
The operator can confirm the ROI by using the ‘Crop Image’ button. The cropped image isdisplayed in the ‘Cropped Image’ section for confirmation with the operator. The operator can alsocrop the image again, in case of improper selection of ROI. Once the cropped image is finalized,the class of hydrophobicity is determined using the ‘Classify’ button. The ‘Classify’ button is used toactivate the feature extraction process and evaluation of ANFIS classifier using these extracted features.The ‘output’ section provides the class of hydrophobicity of the insulator and also remarks about thecondition of the insulator. Thus, the GUI facilities provide easy access to the operator without anyneed to learn programme using the proposed ANFIS classifier.
5. Conclusions
This work proposes a design of intelligent ANFIS to classify hydrophobicity of polymer insulators.Dispersion of water droplets was used as a marker for classification and was evaluated using featuresextracted from its image. A total of 10 statistical image features were used in this study, and the bestfeatures were investigated using PCA. The six best features were selected before the design of theANFIS classifier. The accumulation effect on classification accuracy was evaluated after designing theANFIS classifier. This reduces the number of features required for classification to be four, from the sixbest features identified by PCA.
The four best features namely average intensity, correlation, skewness and homogeneity, werefound to be optimal for the proposed work. Triangular and trapezoidal MF’s with linear andconstant output functions were used for the ANFIS classifier, and their performances were evaluated.Experimental results illustrate a reliable classification accuracy of 94.85% when triangular MF’sand linear functions are employed in ANFIS design. Field tests were also carried out in threeidentified, different sites. Field insulators were to be cleaned before acquiring images and requiredadditional pre-processing to reduce the illumination effects. Field results also exhibit 91.12%, which
Energies 2018, 11, 3391 15 of 16
is a satisfactory level of classification accuracy. A GUI was designed and deployed to facilitate theoperators to use the proposed hydrophobicity classifier.
The proposed procedure can be applied for applications demanding image based classification.Applications such as thermographic inspection of electrical installations, flash over characterizationin contaminated insulators, and electrical fault detection using thermal images can be carried outusing the proposed technique. Use of unmanned aerial vehicles to acquire the images and classifyhydrophobicity online by the proposed classifier can be investigated in future. Incorporation ofenvironmental and pollution conditions on the surface of polymeric insulators in this classifier canmake it robust and can improve classification accuracy.
Author Contributions: R.J. conceived, designed, performed the experiments, analyzed the test results and wrotethe manuscript. K.V. supervised the research work. S.R.K. and R.J. contributed software analysis. All authorsrevised and proofread the manuscript.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflicts of interest.
References
1. Looms, J.S.T. Insulators for High Voltages; IEE Series; IET: Stevenage, UK, 1990.2. Papailiou, K.O.; Schmuck, F. Silicone Composite Insulators: Materials, Design, Applications; Springer: Berlin,
Germany, 2013.3. Wen, X.; Yuan, X.; Lan, L.; Hao, L.; Wang, Y.; Li, S.; Lu, H.; Bao, Z. RTV silicone rubber degradation induced
by temperature cycling. Energies 2017, 10, 1054. [CrossRef]4. Hanada, S.; Miyamoto, M.; Hirai, N.; Yang, L.; Ohki, Y. Experimental investigation of the degradation
mechanism of silicone rubber exposed to heat and gamma rays. High Volt. 2017, 2, 92–101. [CrossRef]5. Mavrikakis, N.C.; Mikropoulos, P.N.; Siderakis, K. Evaluation of field-ageing effects on insulating materials
of composite suspension insulators. IEEE Trans. Dielectr. Electr. Insul. 2017, 24, 490–498. [CrossRef]6. Cheng, L.; Shao, S.; Zhang, S.; Liao, R.; Yang, L.; Guo, C. Research on the long-time operation performance
of composite insulator shed hydrophobicity under hydrothermal conditions. High Volt. 2018, 3, 67–72.[CrossRef]
7. Du, B.X.; Li, Z.L. Hydrophobicity, surface charge and DC flashover characteristics of direct-fluorinated RTVsilicone rubber. IEEE Trans. Dielectr. Electr. Insul. 2015, 22, 934–940. [CrossRef]
8. Hussain, M.M.; Farokhi, S.; McMeekin, S.G.; Farzaneh, M. Risk assessment of failure of outdoor high voltagepolluted insulators under combined stresses near shoreline. Energies 2017, 10, 1661. [CrossRef]
9. Xu, Z.; Lü, F. A static contact angle algorithm and its application to hydrophobicity measurement in siliconerubber corona aging test. IEEE Trans. Dielectr. Electr. Insul. 2013, 20, 1820–1831. [CrossRef]
10. Barsch, R.; Jahn, H.; Lambrecht, J.; Schmuck, F. Test Methods for Polymeric Insulating Materials for OutdoorHV Insulation. IEEE Trans. Dielectr. Electr. Insul. 1999, 6, 668–675. [CrossRef]
11. IECTS 62073: Guidance on the Measurement of Wettability of Insulator Surfaces; International ElectrotechnicalCommission: Geneva, Switzerland, 2003.
12. Berg, M.; Thottappillil, R.; Scuka, V. Hydrophobicity estimation of HV polymeric insulating materials.Development of a digital image processing method. IEEE Trans. Dielectr. Electr. Insul. 2001, 8, 1098–1107.[CrossRef]
13. Zhao, W.; Huang, L.; Wang, J.; Sun, Z. Entropy maximization histogram modification scheme for imageenhancement. IET Image Process. 2014, 3, 226–235.
14. Zhao, W.; Huang, L.; Wang, J.; Sun, Z. Combination of contrast limited adaptive histogram equalisation anddiscrete wavelet transform for image enhancement. IET Image Process. 2015, 9, 908–915. [CrossRef]
15. Thomazini, D.; Gelfuso, M.V.; Altafim, R.A.C. Analysis of entropy and fractal dimension to classify thehydrophobicity in polymeric insulators. In Proceedings of the International Symposium on ElectricalInsulating Materials, Mie, Japan, 7–11 September 2008; pp. 279–282. [CrossRef]
16. Xu, Z. Automatic static contact angle algorithm for blurry drop images and its application in hydrophobicitymeasurement for insulating materials. IET Sci. Meas. Technol. 2015, 9, 113–121. [CrossRef]
Energies 2018, 11, 3391 16 of 16
17. Xu, Z. Static contact angle algorithm selection for superhydrophobic surface hydrophobicity detection.Micro Nano Lett. 2014, 9, 6–10. [CrossRef]
18. Dong, Z.; Fang, Y.; Wang, X.; Zhao, Y.; Wang, Q. Hydrophobicity classification of polymeric insulators basedon embedded methods. Mater. Res. 2015, 18, 127–137. [CrossRef]
19. Jarrar, I.; Assaleh, K.; El-Hag, A.H. Using a pattern recognition-based technique to assess the hydrophobicityclass of silicone rubber materials. IEEE Trans. Dielectr. Electr. Insul. 2014, 21, 2611–2618. [CrossRef]
20. Pylarinos, D.; Lazarou, S.; Marmidis, G.; Pyrgioti, E. Classification of surface condition of polymer coatedinsulators using wavelet transform and neural networks. In Proceedings of the 2007 International Conferenceon Wavelet Analysis and Pattern Recognition, Beijing, China, 2–4 November 2007.
21. Wang, Q.; Zhong, Z.; Wang, X. Design and implementation of insulators material hydrophobicity measuresystem by support vector machine decision tree learning. In Proceedings of the Fourth InternationalConference on Machine Learning and Cybernetics, Guangzhou, China, 18–21 August 2005.
22. Jang, J.S. ANFIS: Adaptive-network-based fuzzy inference system. IEEE Trans. Syst. Man Cybern. 1993, 23,665–685. [CrossRef]
23. Haralick, R.M.; Shanmugam, K.; Dinstein, I.H. Textural features for image classification. IEEE Trans. Syst.Man Cybern. 1973, 3, 610–621. [CrossRef]
24. Thomazini, D.; Gelfuso, M.V.; Altafim, R.A.C. Hydrophobicity classification of polymeric materials based onfractal dimension. Mater. Res. 2008, 11, 415–419. [CrossRef]
25. Hydrophobicity Classification Guide. Available online: https://www.stri.se/wwwpublic/STRI_Guide_1_92_1.pdf (accessed on 12 November 2018).
26. Jolliffe, I.T. Principal Component Analysis; Springer-Verlag: New York, NY, USA, 2006.27. Takagi, T.; Sugeno, M. Fuzzy identification of systems and its applications to modeling and control.
IEEE Trans. Syst. Man Cybern. 1985, 15, 116–132. [CrossRef]28. Rahbari, O.; Mayet, C.; Omar, N.; van Mierlo, J. Battery Aging Prediction Using Input-Time-Delayed Based
on an Adaptive Neuro-Fuzzy Inference System and a Group Method of Data Handling Techniques. Appl. Sci.2018, 8, 1301. [CrossRef]
29. Petkovic, D.; Cojbašic, Ž.; Nikolic, V.; Shamshirband, S.; Kiah, M.L.; Anuar, N.B.; Wahab, A.W. Adaptiveneuro-fuzzy maximal power extraction of wind turbine with continuously variable transmission. Energy2014, 64, 868–874. [CrossRef]
30. Petkovic, D.; Issa, M.; Pavlovic, N.D.; Zentner, L.; Cojbašic, Ž. Adaptive neuro fuzzy controller for adaptivecompliant robotic gripper. Expert Syst. Appl. 2012, 39, 13295–13304. [CrossRef]
31. Méndez, G.M.; De los Angeles Hernandez, M. Hybrid learning for interval type-2 fuzzy logic systems basedon orthogonal least-squares and back-propagation methods. Inf. Sci. 2009, 179, 2146–2157. [CrossRef]
32. Chai, T.; Draxler, R.R. Root mean square error (RMSE) or mean absolute error (MAE)?—Arguments againstavoiding RMSE in the literature. Geosci. Model Dev. 2014, 7, 1247–1250. [CrossRef]
33. Mathur, N.; Glesk, I.; Buis, A. Comparison of adaptive neuro-fuzzy inference system (ANFIS) and Gaussianprocesses for machine learning (GPML) algorithms for the prediction of skin temperature in lower limbprostheses. Med. Eng. Phys. 2016, 38, 1083–1089. [CrossRef] [PubMed]
34. Kaur, R.; Sangal, A.L.; Kumar, K. Modelling and simulation of adaptive Neuro-fuzzy based intelligentsystem for predictive stabilization in structured overlay networks. Eng. Sci. Technol. Int. J. 2017, 20, 310–320.[CrossRef]
© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open accessarticle distributed under the terms and conditions of the Creative Commons Attribution(CC BY) license (http://creativecommons.org/licenses/by/4.0/).