Measurements such as current measurement, voltage measurement, and both current-voltage measurements extract features for fault analysis. An HIF is accompanied by the intermittence of arc (Chen et al., 2013). The arcing fault contains low- and high-frequency components in the current frequency spectrum. The low frequency-based technique results in lower-order harmonics with even, odd, and intermittent harmonics extracted for HIF detection. High frequency-based techniques show short variations in the HIF current.

Voltage measurement is performed by extracting three-phase voltage signals proposed by Ali et al. (2014) during the HIF phenomenon in an underground distribution network. Bakar et al. (2014) performed a voltage measurement at the primary substation and compared fault features with the database generated from the simulation. The method has a single measurement and multiple branches that can detect multiple faulty sections. Detection of HIFs by voltage measurement is efficient only when there is a voltage drop between the relay and fault location. The proposed method by Wang et al. (2018) used the discriminant vector of negative and zero sequence current and voltage in the substation.

Current and voltage measurement has improved reliability compared to the latter measurements. Magnetic field intensity measurement increases the cost and complexity of the detection technique (Bahador et al., 2018).

Signal processing techniques are widely used to improve the effectiveness of high-impedance defect detection approaches. Signal processing techniques extract the hidden properties of observed three-phase signals for HIF detection, enhancing adaptability, stability, and cost-effectiveness. More informative data are obtained using various analyses based on these extracted data, such as time-domain analysis, frequency-domain analysis, and time–frequency analysis (Chen et al., 1990; Sarlak and Shahrtash, 2013). Table 1 gives a detailed comparison of various signal processing techniques for HIF detection using intelligent classifiers.

Shahrtash and Sarlak (2006) used a pattern recognition-based approach for HIF detection with DT as the classifier. The power distribution system is illustrated in which the system voltage is 12.5 kV, the short circuit level (at the infinite bus) is 866 MVA, and a time constant of 45 ms is shown in Figure 4A. The following data about transmission lines are given inductance of transmission line of 825 nH/m, resistance of transmission of 313 Ω/m, and line length of 33 km. The loads connected are a capacitor load rated 4.08 MVAr, transformer (10/0.4 kV) connected in delta-star, three-phase thyristor converter as harmonic load, and nominal load current of 630 A. The best results are obtained in even, odd, and in-between harmonics below 400 Hz. The classification factor is based on entropy, which is the effectiveness of an attribute in classifying data. A total of 2,583 and 1,331 cases were used for training and testing purposes, respectively. For different values of the sampling frequency of 2 kHz, a data window size of 2 cycles, and a pre-processing-time interval of 30 cycles, the accuracy of the proposed method was 99.4%.

Vahidi et al. (2010) used the DWT technique to extract the features, and ANN classifies the faulty cases with other power system disturbances. A three-phase radial distribution network is modeled using the PSCAD/EMTDC software used as the test system. The power system frequency is 50 Hz, and the power is supplied at 63 kV from a 30-MVA transformer (wye/delta). The transformers and line parameters are shown in Figure 4C. Line currents during the HIF have high-frequency components and are used for feature extraction. The extracted features are decomposed into two levels of detailed and approximation coefficients at six cycles. The performance of the DWT-based denoising technique depends on the threshold value γ , which can be divided into hard thresholding and soft thresholding. A large value of γ will shrink most of the coefficients to zero and for small values, denoising outcomes are inefficient. The extracted data are trained and compared with six different types of mother wavelet transforms: haar, coif2, dmey, db9, bior2.6, and sym8. Sym8 gives the best performance and accuracy. The sampling rate chosen is 20 kHz. A multi-layer perceptron neural network (MLPNN) with a Levenberg–Marquardt algorithm is used. The effect of SNR on the proposed algorithm is also studied. The HIF model is introduced to buses 2 and 3 to mimic the HIFs that exist on different ground conditions, such as sand, wet soil, dry soil, asphalt, and grass, giving an accuracy of 99%.

The system tested by Tawafan et al. (2012) is a 115-kV distribution feeder comprised of a substation, and three radial network distribution feeders are shown in Figure 4B. The generator is 30 kV and 10 MV connected to the 30/13.8-kV and 10-MV transformers. The 6-pulse rectifier is used for the representation of the non-linear load. The simulation models are created using PSCAD, and the sampling rate is 15.36 kHz. FFT is the feature extraction technique used, with an algorithm based on the adaptive neural Takagi–Sugeno–Kang (TSK) fuzzy modeling scheme, where the HIF detection is performed by taking the amplitude of the ratio of the second and odd harmonics to fundamental harmonics of the current signals that serve as input to ANFIS. The fundamental harmonics are decreased when the fault has occurred. A total of 570 cases are taken, among which 138 cases are HIFs and 432 are non-HIFs. The mean squared error value of the model is 0.1163, and based on the output of ANFIS, if the HIF current is greater than 0.6 it indicates HIF conditions and if it is less than 0.4; it is non-HIF conditions. The detection accuracy of HIF cases is 97.8%, and non-HIFs is 99%. The method was proposed by Soheili et al. (2018). The harmonic components of the third, fifth, seventh, eleventh, and thirteenth HIF current are preprocessed in an EKF, and 12 features are extracted. These features of one-, two-, and three-cycle windows are considered the input to train the RF. RF is trained with 20,580 and 8,820 data sets. The SNR chosen is 20 dB. Two separate, three-phase sources are connected through transformers to a transmission line of length 100 km. The transmission lines are 138 kV, and the transformers are 50 MVA, supplying at 138/25 kV to the distribution network (Figure 4D).

The distribution feeders (pi sections of 20 km each) work at 25 kV and are connected with shunt capacitors, linear loads, and a 2-MVA 6-pulse rectifier load (non-linear load). The resistance, inductance, and capacitance of positive and zero sequences of transmission lines are as follows: R₁ = 0.01273 Ω/km, X₁ = 0.9337 mH/km, C₁ = 0.0012 lF/km and R₀ = 0.3864 Ω/km, X₀ = 4.1264 mH/km, and C₀ = 0.0075 lF/km, respectively. The resistance, inductance, and capacitance of distribution lines (pi-section) are R₁ = 0.2568 Ω/km, X₁ = 2.0 mH/km, and C₁ = 0.0086, respectively. The total percentage impedance of the transformers is 6.75%. The simulation models are developed using PSCAD (EMTDC), and the sampling rate chosen is 1.0 kHz on a 50-Hz base frequency (20 samples per cycle). RF proved to be the best, with 93.56% accuracy and 94.56% dependability. A multi-feeder radial distribution system was proposed by Sarwagya et al. (2018) to detect and segregate HIFs. It consists of a 30-MVA, 33-kV substation, and five numbers of 11 kV radial distribution feeders. The positive-sequence impedance of the distribution line is 0.3 + j0.25 Ω/km. The discrimination of the HIF is performed based on two criteria. The first is based on the maximum value of the one-cycle sum of superimposed components of negative-sequence current for faulty feeder identification, and the second is based on the one-cycle sum of superimposed components of residual voltage for HIF detection. The substation bus provides the residual voltage. The negative sequence current of all the feeders is compared, and the maximum value of the negative sequence will be for the feeder where HIF has occurred. The HIF is accurately detected in 3 s with the proposed method. The performance of the method with HIF during unbalanced loading, unbalanced loading conditions, capacitor switching, and occurrence of an HIF in various feeders is analyzed. Sarlak and Shahrtash (2011) compared two approaches for HIF identification: the voltage phase difference algorithm and a combination of DWT and ANN. The test system chosen is from Bosnia and Herzegovina, a distribution system like in Europe. The article includes two test systems: a two-bus feeder system and the other is an eight-bus system. The two-bus system is a simple one with a main transformer of 35/10 kV, whereas the latter one is more complex, consisting of 8 feeders fed from 35/10 kV, with underground cables and a transformer at the end consumers rating at 10/0.4 kV. Figures 4E, G represent two- and eight-bus systems, respectively. The sampling frequency is 3.2 kHz. The first method DWT is applied to the measured voltage signals. Each voltage has four detailed coefficients and one approximation coefficient. An algorithm is proposed in which DWT signals are combined, representing a signature for symmetrical and unsymmetrical faults. These data are then used to train and test the ANN. A total of 1,600 cases are simulated, including non-faulty conditions and three types of fault conditions. The method gives an accuracy of 100% for the 20–600 Ω range of fault resistances and at different fault locations. In the second method, a voltage measurement is performed, and the Hilbert transform is applied to obtain the best features. The best feature is an instantaneous frequency, which represents the time rate of change of the instantaneous phase angle. The phase difference is calculated by the difference between the instantaneous phases of voltage signals. The voltage phase difference algorithm calculates the PD during normal and fault conditions. At normal working operation, the phase difference will be 120⁰, and during each fault condition, the phase difference will be different. This parameter is used to detect and classify the fault. With 2,000 cases in two-bus systems, the accuracy obtained is 92.5%, whereas the eight-bus system showed 90.19%. Panigrahi et al. (2018) used the IEEE Power System Relaying Committee Working Group at medium voltage levels. A simple 11-kV radial distribution feeder with eight nodes is shown in Figure 4F. MATLAB/SIMULINK program is selected and modeled, and the line impedance (positive sequence) is chosen as 0.3 Ω/km + j 0.25 Ω/km. At a distance of 5 km, nodes are isolated from each other, in which nodes 1, 3, 5, and 7 are connected to linear capacity loads of 1 MVA each at power factor 0.9/phase, and nodes 2, 4, 6, and 8 are linked to linear capacity loads of 2 MVA each at power factor 0.9/phase. The method discussed the MM gradient for HIF detection and classified HIF, LIF, capacitor switching, and load switching (balanced and unbalanced).

The proposed method measures the three-phase voltage at a relay location and evaluates the residual voltage. The morphology gradient is used to extract the irregularities in the voltage signal. The extracted feature index is determined from the zero-energy index at NC and compared with a predefined threshold value. The extracted feature index value will jump slightly for HIF, LIF, and other disturbances for faulty conditions. An HIF is created at nodes 1, 4, and 9, and the occurrence will be for 1 s. The method accurately detects HIF occurrence in 0.8 s. The test system model proposed by Wali et al. (2018) is a 13.8-kV radial distribution feeder simulated by MATLAB/SIMULINK under different scenarios, such as linear load, non-linear load, and several other conditions. Figure 4H shows a single-line representation, a three-phase transformer, and a 13.8-kV distribution network. The non-linear load is represented by a 6-pulse rectifier that creates non-linear features in the feeder. The method used for HIF detection is FFT for feature extraction, and the power spectrum technique is used to identify the fault, which gives an accuracy rate of 100%, the time required is less, and that does not require any level of training. FFT extracts the feature of the current signal from the faulty feeder, and the power spectrum of the time signal is determined using the function FFT. If PS is less than 0.005, then a HIF occurs. HIF of 250 cases and other power system disturbances of 750 cases have been analyzed in this study. The method distinguishes events due to capacitor banks, non-linear loads, linear loads, and HIF.

The test system used for HIF detection (Veerasamy et al., 2018) consists of a grid source of 50 MVA/30 kV, a distribution transformer (12 MVA, 30 kV/13.8 kV), a common bus of 13.8 kV, and five radial type distribution feeders, integrated into the load facility. An Emanuel two-diode model consisting of two variable DC voltage sources of 1–10 kV connected to anti-parallel diodes by non-linear resistors of 50–500 Ω is considered an HIF model with non-linear arc characteristics. The method is proposed by extracting the features using CWT and DWT and classifying the extracted features by ANFIS. CWT gives the region at which the fault has occurred, and DWT can locate it by calculating the standard deviation (SD) using a five-level decomposition. The extracted SD values of different fault conditions with different values of fault resistance from the detailed and approximation coefficients are obtained, which are used to train classifiers FLS and ANFIS. Various faults, such as symmetrical, unsymmetrical, and HIFs, were tested using MATLAB/SIMULINK. The classification rate of ANFIS is 100%, which proved more effective than FLS. Wang et al. (2018) proposed that an HIF detection algorithm identifies the non-linear voltage–current characteristic profiles (VCCP) for identifying an HIF in the MV distribution system. During HIF, the zero-sequence current is less than the positive-sequence current. The slope of the VCCP is the numerical difference between voltage data from current sample data, and the least square linear fitting method is proposed. The wavelet correlation coefficient (WCC) is considered to improve the reliability of the algorithm. If WCC is greater than 0.966, the metered data are from a faulty feeder, and if less than the value, it is a healthy feeder. The radial distribution system is the test system in Figure 4J that uses EMTP/ATP program. The typical Mayr arc model is simulated and drawn in series with constant resistance using a switcher and parallel branches. The simulation time stage was set at 2 μs field-metered data from KEPCo, South Korea, and HIF experiments were performed on a 22.9-kV no-load overhead feeder to check the simulations. As it is a no-load feeder, the zero-sequence current and phase voltage were used at a faulty feeder outlet to correctly estimate the fault point voltage and fault branch current. The faults have been tested on various dry surfaces. The algorithm showed excellent results in real-time digital simulator tests. The test system proposed by Sedighi et al. (2005b) for HIF tests and data collection is a radial feeder of 20 kV at Qeshm Island, Iran, as shown in Figure 4K. The feeder is energized from another 20-kV feeder by two distribution transformers (20/0.4 kV, 100 kV A) connected back-to-back. The HV and LV connections of the transformers are delta/star connected. The HV sides of the transformers are connected to feeders, and the LV sides are connected to the low-voltage switch. Three-phase voltages and currents were monitored and recorded using Hall effect current transformers, potential transformers (PT), power analyzers, and computers. The sampling rate of the recorded data was 24,670 kHz for each test, and the overall recorded time was 15 s. The method used for HIF detection uses WT for feature extraction with a three-level decomposition of current signals. The first method uses GA for feature vector reduction, and the Bayes classifier is used for classification. In the first method, coefficients of three-level decomposition are used for feature extraction. They are divided into 10, 5, and 5 segments. In GA, each segment is mapped to a 20-dimensional space. A space with 20 dimensions is mapped to a space with five dimensions. The Bayes classifier is used to classify the mapped space. In the second method, WT transforms are also applied for the current signals, PCA is used for feature vector reduction, and NN is the classifier. The coefficients of three-level decomposition are used for feature extraction, divided into 10, 5, and 5 segments. Means of the absolute value of each segment were chosen as features, and the extracted signals were mapped to a 20-dimensional space. Using PCA, space was reduced to a 7-dimensional space. A perceptron NN using backpropagation discriminates between HIF, isolator leakage current, and other power system transients. Sekar and Mohanty (2020) stated that morphology gradient extracts the features of which a rule is set by RF and then fed to a fuzzy rule-based algorithm for HIF detection. The electric power distribution system (EPDS) was modeled using MATLAB/SIMULINK. A three-phase shunt capacitor of 1 Mvar has been connected to the busbar to improve power quality and output. The inrush current of a transformer produces an asymmetrical current signal that may serve as a transient signal generated by switching that may be like the HIF current waveform. An induction motor is connected as a load to study the motor operation impacts. The EPDS uses linear and non-linear loads to simulate the loading scenario, as shown in Figure 4I. One cycle window length of the current signal is measured, the impulsive feature of the signal is extracted, and a rule set is created from the statistical features of RF. The sampling frequency is 1,000 Hz, and the signal length is 0.5 s. The periodic signal has third- and fifth-order harmonics, and other harmonics that are negligible make the filter closer to HIF detection. The method effectively detects an HIF from other power system disturbances.

The test system proposed by Lai et al. (2005) consists of two 50 MVA generators with 25 kV lines, two transformers, and linear and non-linear loads. In DWT, db4 was chosen as the mother wavelet, with a downsampling frequency chosen as 9,600 Hz, used to extract the detailed and approximation coefficients from the signals of the HIF and non-fault. The current and voltage signals at a targeted circuit breaker are measured. The RMS values of the measured quantities at various frequencies are analyzed and given as input to the nearest neighbor to classify fault signals. The HIF and non-HIF cases (1,000 cases each) were simulated with HIF and LIF models. The range of total error corresponding to the RMS value of the voltage wavelet coefficient is from 2.52% to 45.4% and will not indicate the physical properties of the output coefficient.

Samantaray et al. (2008) reported that S transform and TT extract the features, and FNN and PNN classify the faulty and non-faulty conditions of the HIF. The ST features are extracted from the HIF and normal fault current signals for half-cycle current signals after fault inception. The energy and SD of time and frequency information are considered feature sets. These features are used to train and test the FNN for radial and mesh networks. TT transform also extracts energy and SD of the TT-counter and time index after fault inception for the first half-cycle of the fault current. A total of 500 cases are simulated for training and testing the classifier. The system is modeled in MATLAB/SIMULINK, and the sampling rate chosen is 1 kHz. The PNN classification is based on the distribution values of the probability density function. The classification rate of a radial network with PNN using ST features is up to 98.06%. PNN with TT transform features gives a classification rate of up to 98.05%. The accuracy of the FNN and S transform combination is 93.04%, and the accuracy of the TT transform is 94.16%. The proposed method is also tested in a mesh network. The accuracy of the mesh network with S transform and FNN is 92.86%, and that with TT transform and FNN is 93.55%. The S transform and PNN combination gives 97.85% accuracy, and the TT transform and PNN combination gives 97.09% accuracy.

The HIF detection method used by Samantaray et al. (2009a) is a combination of AEKF with FNN and PNN. The schematic diagram of the test system chosen is given in Figure 5. The base voltage of the distribution network is 25 kV, and the generator is 10 MVA, 15 kV capacity. The harmonic components estimated by the AEKF are fundamental, third, fifth, seventh, eleventh, and thirteenth harmonics for the HIF and NF under non-linear loads. The AEKF calculates the harmonic component within half a cycle of the fault occurrence, with the peak of the estimated harmonic component considered that inputs to PNN and FNN. The PNN classification is based on the probability density function’s distribution values. PNN is analyzed using a data set with an SNR of 20 dB, 300 data sets for training, and 200 data sets for testing. For the classification, PNN takes 0.1 s time, whereas FNN takes 0.2 s. Fault and non-fault conditions with non-linear switching (a six-pulse rectifier is used) are checked using various models of MATLAB/SIMULINK, and the sampling rate chosen is 1.6 kHz. The accuracy rate of PNN is 99.11% compared with FNN having 96.51%.

The detection of HIF described by Routray et al. (2016) uses a test system with a generator of 50 MVA supplying 138 KV of voltage to the utility sector through a transmission line 100 km long, and a 138/25-kV star/delta transformer is considered for testing the method. The method uses ST for feature extraction and ANN for discriminating the HIF with load switching, capacitor switching, and NC. The time and frequency information is extracted from the S matrix, and the amplitude factor is calculated from current signals. A total of 4,010 cases were considered, of which 60% is used for training and 40% for testing. The overall accuracy of classifiers for normal fault is 98.75%, 96.4%, 94.06%, and 92.60% for normal (without noise) and noisy conditions.

Samantaray (2012) studies two test systems: one with a radial feeder mentioned in Figure 5 and the other with a mesh feeder given in Figure 7C. The test system studied is connected to a 50-MVA generator and a transformer of 138/25 kV from a transmission line of 138 kV and a length of 100 km. Loads are connected with linear and non-linear loads. The resistance, inductance, and capacitance of positive and zero sequences of transmission lines are R₁ = 0.01273 Ω/km, X₁ = 0.9337 mH/km, C₁ = 0.0012 lF/km and R₀ = 0.3864 Ω/km, X₀ = 4.1264 mH/km, and C₀ = 0.0075 lF/km, respectively. The resistance, inductance, and capacitance of distribution lines (pi-section) are R₁ = 0.2568 Ω/km, X₁ = 2.0 mH/km, and C₁ = 0.0086, respectively. The total percentage impedance of the transformers is 6.75%. The simulation models are developed using PSCAD (EMTDC), and the sampling rate chosen is 1.0 kHz on a 50-Hz base frequency (20 samples per cycle). RF proved to be the best, with 93.56% accuracy and 94.56% dependability. On the distribution feeder, the HIF faults are generated at 25 kV, 20 km, pi section. Different simulation conditions are also considered, such as three-phase loadings, single-phase loadings, transformer energizations, shunt capacitor switching, and HIF by varying DC voltage sources. Two combinations of the HIF detection technique are proposed: the first is an EKF and RF and the latter is DT with FFT. The harmonic components of the third, fifth, seventh, eleventh, and thirteenth HIF current are preprocessed in an EKF, and 12 features are extracted. These features of one-, two-, and three-cycle windows are considered in this work. Considering the two-cycle window and with SNR set at 20 db. The simulation models are developed using PSCAD (EMTDC), and the sampling rate chosen is 1.0 kHz. RF is trained with 20,580 data sets, and 8,820 data sets are tested. RF proved to be the best, with 93.56% accuracy and 94.56% dependability compared with DT.

Mishra et al. (2016) used S transform with ANN and SVM to discriminate the HIF from other power system disturbances. A total of 4,000 cases, including HIF, normal, load switching, capacitor switching, and normal faults, are taken. Features are extracted from three-phase currents measured from the bus, and the best feature vector is selected. MLPNN with backpropagation NN and SVM along with ST is used, with 60% data for training and 40% for testing. The distribution model with a radial pattern of a 50-MVA generator is connected to a 100-km-long transmission line and a 138/25-kV star/delta transformer to supply 138 kV voltage to the utility sector. For the ideal case, ST with ANN and ST with SVM give 100% accuracy, whereas with SNR 30 dB, ST with ANN gives 93.7% and ST with SVM gives 92.15% accuracy. For the ideal case with the mesh network, ST with ANN and ST with SVM give 100% accuracy, whereas with SNR 30 dB, ST with ANN gives 81% and ST with SVM gives 86% accuracy.

Detection of the HIF using a combination of MLPNN based on multi-resolution morphological gradient features of the current waveform is described by Sarlak and Shahrtash (2011). The MMG features of the current signals (for three half-cycles) of broken and unbroken conductors are considered, and the features from DFT, DTT, DST, and DWT are compared. The morphology gradient is the difference between the dilation and erosion functions. Data acquisition is performed using the ION7650 meter, and the input port of the meters is connected to the outputs of the current transformers at the 63/20-kV substation. The sampling rate of the current waveform is 1.6 kHz. A disturbance detection module is based on MMG-extracted features of any subwindow with a predefined threshold. Three MLPNNs (A, B, and C) are trained individually by applying the time-based features obtained from the first, second, and third sub-windows. Then, their decisions are concatenated to make the final decision. The proposed algorithm gives security of 96.3% and dependability of 98.3%.

Nikoofekr et al. (2013) used a test system from Tehran, Iran, which has a 63/20-kV transformer feeder with 30 MVA apparent power, and the HV side has been grounded with a zigzag transformer and variable resistance adjusted at 29.5 Ω. Moreover, two 2.4-MVAR capacitor banks are connected through the HV circuit breakers. The ION 7,650-meter tests the HIF current and non-HIF current signals, such as insulator leakage current (ILC) and harmonic load current, with a sampling rate of 64 samples per cycle at the site. The method uses ST for phase correction in CWT, which localizes the phase and amplitude spectrum. TT transform extracts the features of the measured signals. Five different ART neural networks are used to classify the HIF and tested with broken conductors in asphalt, concrete, gravel surfaces and unbroken conductor on trees and under no fault conditions. This study uses five types of ART networks, namely, ART1, ART2, ART2-A, Fuzzy ART, and Fuzzy ARTMAP. The different features extracted were energy, SD, and median absolute deviation. The performance of the ART network is based on the vigilance parameter ρ whose value ranges from 0 to 1. For lower values of ρ , the classification is rough, and categories are less, and for higher values of ρ the categories are more, and classification is fine. In the basic ART training process, the input pattern is delivered to the input layer, which activates F2 neurons via bottom-up weights. Because the F2 layer is a competitive layer, the neurons compete with each other to learn the input vector, and the larger neuron wins. All other F2 units’ activations (outputs) are set to zero. Then, the top-down weights of the winner neuron are sent back to the F1 layer. Figure 6 shows the Palash feeder in Tehran. From the total 6,437 data taken, 60% were used for training and 40% for testing. The result of the network shows an accuracy of ART1 of 91.61%, ART2 of 98.65%, ART2-A of 99%, Fuzzy ART of 99.18%, and Fuzzy ARTMAP of 99.18%.

MMG is the feature extraction used by Sarlak and Shahrtash (2013) and tested on the Palash feeder in the Southwestern Tehran distribution network, as shown in Figure 6. An HIF indicator is installed in various poles that detect HIF at various locations. The HIF indicators are installed in the feeder based on the processing of the magnetic-field strength signal. The fitness evaluation combines three goals: accuracy, number of training samples, and the weighting factor. The impulse response of magnetic response in the frequency domain is calculated in terms of the electric hertz vector. By simulation, a 978-feature vector for the HIF and 852 non-HIF is calculated. The dependability and security of the proposed system are best above the 20 db SNR. To evaluate the proposed method, MMG extracts a magnetic field strength signal, which is given to SVM for classification. The proposed algorithm has 96.9% security and 97.2% dependability.

The real-time experiments are performed in the Palash feeder, Tehran (Soheili et al., 2018). The modified FFT approach is used to detect the HIF concerning non-linear loads. In the proposed method, the measured three-phase current is analyzed by FFT. These currents are continuously monitored for non-linear loads, abnormal conditions, and HIF detection. HIF currents during non-linear loading conditions and different ground types are recorded. The proposed algorithm has divided the output into three levels: 0, 0.5, and 1. NC, pickup, and HIF, respectively, are represented by these levels. Various scenarios in the simulated data are considered, including high current three-phase feeder, low current three-phase feeder, low current single-phase feeder, and capacitor switching events. The distribution network is energized via a 63-kV/20-kV three-phase transformer with a rated power of 30 MVA. Data recording has been conducted using the ION 7650, with a sampling rate of 64 samples per cycle (3.2 kHz). The various surfaces where real-time experiments are conducted are concrete with 20 cm in no-load conditions, concrete with 10 cm in 55% full load conditions, and asphalt with 2 cm under 55% full load conditions. The HIF detection time of the proposed method was 1.13 s.

There are various IEEE test systems, such as the single-line diagram of the IEEE 13-node system (Figure 7A), the single-line representation of the IEEE four-node test system (Figure 7B), the IEEE-123 distribution feeder (Figure 7C), and the IEEE 34-node test system (Figure 7D).

The test microgrid system is used for HIF detection (Kar and Samantaray, 2017), as shown in Figure 8A. The base power of the test system is chosen as 10 MVA. The rated short-circuit of the utility is 1,000 MVA with f = 60 Hz, rated 120 kV. Distribution generations, DG1 and DG3, are rated as follows: synchronous generator rated at 9 MW and rated voltage of 2.4 kV, and DG2 is a wind farm consisting of three wind turbines (2 MW each), rated kV = 575 V. The transformer ratings used in this study are as follows: Transformer 1: 15 MVA, 120/25 kV. Transformers 2 and 4 are rated at 12, kV = 2.4 kV/25 kV, while Transformer 3 is rated at 2.5, kV = 575 V/25 kV. The distribution lines (DL) are DL1, DL2, DL3, and DL4: PI-Section, 20 km each. The total load is 20 MW, 10 MVAR, a sum of L1–L5. The MODWT is the feature extraction technique, and DT is the classifier used. The proposed method is tested in both grid-connected and islanded modes. The MODWT scaling filter and the wavelet filter related to the DWT filter are calculated, and the scaling coefficients of MODWT are obtained. The detailed approximation coefficient is obtained from the MODWT, and DT does accurate classification. The total cases are 1,493, of which 973 are HIF cases and 520 are faulty conditions. In the method proposed, 12 feature sets are considered, among which five were taken for classification. The training set (70%) and testing (30%) assess the performance. The software package “R” generates data mining for the DT. The detection accuracy, dependability, and security are 100% for the grid-connected mode, whereas for the islanded mode, the accuracy is 99.23%, security is 98.23%, and dependability is 100%.

Microgrid is considered while integrating distributed energy systems (Abdelgayed et al., 2017). The Consortium for Electric Reliability Technology Solutions (CERTS) was used for the case study of microgrids in this article. The microgrid system has two modes of operation: grid-connected and islanded mode of operation. CERTS microgrid consists of a distribution system fed from three-phase distribution transformers rated at 13.8/0.48 kV, consisting of two solar photovoltaic sources and one battery energy storage source. Four loads are considered long in the distribution system. The method employs a semi-supervised machine learning strategy to handle labeled and unlabeled data. DWT extracts the hidden properties of voltage and current and applies them to a harmony search algorithm to find the HIF parameters. The DT and KNN classifiers are used to discriminate the HIF events. The overall accuracy of the DT is 100%, and that of the KNN is 95%.

The test system is a CERTS microgrid with two inverter-interfaced DG units and one synchronous generator-based DG unit (Gadanayak and Mallick, 2019). The representation CERT microgrid system is shown in Figure 8B. For HIF detection, the test system consists of five distribution lines and five relay units. MATLAB-SIMULINK is used to simulate the model, with a simulation sampling rate of 0.5 MHz. The MODWT approach for feature extraction and knot-based empirical mode decomposition is included in the methodology. The program recognized 855 cases of HIFs and 801 cases of non-HIFs. The average time to detect a fault was 0.12175 s. The test system used by Gashteroodkhani et al. (2020) was performed in a 25-kV IEC standard microgrid that gave high accuracy and robustness in noisy environments. The single-line representation of the IEC microgrid is represented in Figure 8C. A deep-belief neural network with TT-transform is employed where an intelligent relaying scheme-based real-time digital simulator is used, integrated with MATLAB. The process involves the measurement of three-phase currents at both ends and the feature extraction by Clark’s transformation and TTT, which is sent to the DBNN. Six features are used for feature extraction, including energy, SD, and median absolute deviation. Microgrid models with grid-connected, islanded, radial, and mesh topologies are used to test the approach. With 3,600 fault situations and 3,125 no-fault cases, the sampling rate was set to 1.2 kHz. The proposed method gives 99.74% and 99.46% accuracy for a radial network with grid-connected and islanded modes, respectively, and 100% for mesh topology in both modes of operation.

Eldin et al. (2007) considered the JMARTY model with Egyptian transmission line parameters, such as a transmission line length of 125 km and a resistive load of 600 MW, as shown in Figure 9A. DWT’s feature extraction technique with classifier moving window pattern recognition is used in HIF detection in extra-high voltage transmission. The sampling rate is 250 samples/cycle at 50 Hz. In order to distinguish HIF from non-HIF events, the proposed technique uses high-frequency information from wavelet analysis db4. The effect of fault location, fault interception angle, fault type, switching of loads, switching of the HIF, and sudden load rejection is studied. The algorithm can be added to the existing digital relay microprocessor; it is fast, accurate, and simple.

Ibrahim et al. (2008) described two approaches: the first uses DWT analysis and the second analyses three-phase voltages using a high-frequency tap coupling capacitor voltage transformer. The chosen sampling rate is 20,000 Hz. A 100-m-long 345-kV double-end transmission line system is considered. The effect of fault location, fault type, fault interception angle, switching of loads, and switching of HIF are studied. The method is independent of load variations and unbalanced conditions. The algorithm is fast, accurate, and simple; it can be added to an existing digital relay microprocessor (Eldin et al., 2007).

In Yang et al.’s method (2006), the HIF was tested in a real-time Tai-16 feeder, Taishi substation near Mailiao, as shown in Figure 9B. The fault was staged at pole #61 under grounded and ungrounded conditions, with dry and wet ground conditions. Data from the Shand-Ding substation and the She-Zi substation were considered to train the NN. The analog-to-digital recorder ADX-3000 was installed to monitor the neutral line current of the faulted feeder in the substation. A cross-linked polyethylene-covered conductor, a bare copper conductor, and ACSR conductors are used. A wavelet transform and pattern recognition-based feature extraction scheme is used and applied to backpropagation ANN for HIF detection. The detection of the HIF is carried out by introducing an intelligent HIF detector applied to a neutral line current, giving an 81% detection rate. A self-tuning algorithm based on the chi-square algorithm is applied to find the variations in neural current. A statistical confidence scheme is applied for neutral current estimation. For calculating the threshold of neutral current, a 95% confidence interval is applied.

A typical 11-kV distribution network in Malaysia is composed of 38 nodes serving 34-line sections, as shown in Figure 9C, displaying the single-line diagram of a 132/11-kV distribution network. The test unit has a frequency of 50 Hz, with a sampling frequency of 6.4 kHz, and delivers 128 samples/cycle. The feeder bus is simulated using the PSCAD program. Ali et al. (2014) proposed that DWT-based MRA is the feature extraction technique, and a matching approach technique is used for classification in the underground distribution system. The simulation tests of the three-phase voltage signal are obtained at a measurement point using the DWT-based multi-resolution technique. The voltage signal was decomposed into 128 samples and analyzed by Daubechies fourth order of DWT. The first level of the detailed coefficient, d1, detects the HIF by observing changes in the characteristics taken from three-phase voltage data. The sum of the first-level approximation coefficients obtained from the normal cycle, known as the approximation ratio, is divided by the sum of first-level approximation coefficients obtained from the normal cycle to classify the fault: A p p r o x i m a t i o n   R a t i o = ∑ a H I F ∑ a n o r m a l . (6)

The approximation ratio is considered for distinguishing SLGF, LLGF, LLLF, and LLF. If the ratio is less than 1 for one phase and greater for the other two phases, then an SLG fault has occurred. The three-phase ratios will be the same for LLLF. For locating the HIF, a matching technique based on the shortest distance is used. For each line section, the SD between the measured signal and the line section is computed. The summation of the detailed coefficients of voltage signals is done in three coordinate systems and represented in three-dimensional space. The SD of each section is calculated, and the average value is calculated. The average value is compared with the SD values of each section. Each section is then arranged in ascending order of its SD values, which list the possibility of a faulty section. The method successfully located the fault after four or five iterations.

In Cui et al. (2017), the test system chosen is a benchmark system in a remote Canadian community with 25 kV 60 Hz, distribution feeders that serve an 11-MW load built by McGill Power Laboratory, as shown in Figure 9D. An L-G HIF at t = 0.3 s is implemented in a hybrid distribution system. The feeder’s PCC (CB-1) will collect all the data where the HIF detector is located, two DGs at locations A and B, respectively. Three types of distribution systems are also possible to model: 1) the synchronous generator (SG) system is constructed by connecting the SG to location A without DG in location B; 2) the inverter-interfaced system is connected to location B with only type 4 wind turbines at location A and without DG; and 3) the hybrid DG system has an SG connected to location A and wind turbine generators at location B. The rated voltage is 4.17 kV for the SG, and 9 MVA is the power. Three wind turbine generators provide wind power, each rated at 575 V and 2.2 MVA. The proposed method was tested in hybrid systems, and inverter-based systems are grounded, as well as ungrounded conditions. HIF and non-HIF events of 1,944 cases each have been considered for testing and training. The method proposes an algorithm to rank the effective feature sets using a signal processing technique of DFT and Kalman filtering estimation. The effective feature set is derived from information gain or entropy. The information gain of each feature is calculated, and a calculation variable is obtained. Such calculated variables are compared and then ranked for further assessment in HIF detection. By measuring the currents and voltages at the point of the common coupling, 246 electrical features are obtained. Classifiers, such as Native Bayer’s, SVM (Gaussian kernel), k-nearest neighbor (KNN), RF, and J48, are compared, among which RF and J48 proved to be the best with 99% accuracy for both.

Researchers provide numerous HIF models because of stochastic behavior and complicated properties. In order to acquire a good representation of the HIF, it is necessary to develop a model that specifies the features and the harmonic content of the HIF. Because arcing, which has yet to be fully modeled, is involved in most HIF events, an HIF is a difficult example to model. HIFs are non-linear and asymmetric, according to some earlier studies, and random and dynamic arcing features should be used in modeling (Zamanan and Sykulski, 2014). The arc is a continuous luminous discharge of electricity in which many free electrons and ions in an insulating medium are converted into a conducting medium. The arc was first studied as a continuous luminous discharge of electricity through an insulating medium that becomes a conducting medium due to the presence of a large number of free electrons and ions. The arc was first researched about circuit breaker disruption capabilities, with arc models used to improve circuit breaker testing (Elkalashy, 2007). The Emanuel model replicates zero periods of arcing and asymmetry by connecting two DC sources anti-parallel with two diodes (see Figure 10). Variable resistors vary the fault resistance, and a voltage supply with random values mimics HIF unpredictability. Some of the other HIF models are, two-time varying series resistors with different characteristics (Nam et al., 2001), Kizilcay’s model (Zhang et al., 2016; Mishra and Panigrahi, 2019) and Matthews arc model (Gammon and Matthews, 2001).

NN (Snider and Yuen, 1998) models are grouped according to their architecture (Sultan et al., 1992) (which gives the neural connection), processing (describes the production of output corresponding to weight and input), and training (explains the adaption of NN weight for every training vector). The architecture consists of the input, hidden, and output layers. The processed information of the NN is obtained at the output layer. NN architecture types are single-layer networks, such as an MLPNN (Sarlak and Shahrtash, 2011), a Hopfield network, and a Kohonen network (Ebron et al., 1990). ANN implementation is easy but is subject to the amount and quality of trained data. The ANN algorithm needs to be re-trained when there is a change in the data set, and the number of neurons and learning rate are found by the trial-and-error method.

An NN using a relay mechanism with HIF detection is described by Sharaf et al. (1993). The approach of a feed-forward network with backpropagation of one hidden layer and 15 neurons is used, which employs the discriminant vector of negative and zero sequence current and voltage in the substation. Twenty-two cases were taken for training and 10 cases for validation. They are noise-tolerant, require less detection time, and are economical. NN algorithm-based relaying scheme, implemented using NN hardware chips or software, promptly detects HIFs using Fourier analysis of lower-order harmonic vectors of measurements used as the input to the perceptron feed-forward network (Snider and Shan, 1998). The Emanual arc model is used as the HIF model. Logsigmoid hidden layers of three numbers with 10 neurons make the structure. As only lower-order harmonics are utilized for detection, the scheme is more feasible and flexible and has a high detection rate. The scheme is tested using simulation and real-time field measurements. A microprocessor-based pattern recognition technique is developed (Al-Dabbagh and Al-Dabbagh, 1999), which uses DFT to analyze signals. The scheme describes a sensitive Earth fault protection that is comparatively slow but gives better performance. The scheme is not tested with real-time data; only relaying current and voltage signals from the ATP simulation package are tested. Keyhani et al. (2001) used a subband decomposition method for current, which uses the energy of the subband to feed the input vector to the NN. The system is less noise-sensitive and can detect HIFs efficiently at high noise levels. The SNR chosen was 0–14 db. Emanuel and Gulachenski’s HIF model is used for testing. A total of 800 cases were taken, and 16 neurons were used for fault analysis. Two NNs, namely, perceptron and FNN, are used; both give similar results with close to 100% accuracy. The scaling and translation characteristics of DWT are used to discriminate the transient and stable features of current signals (Yang et al., 2004). The extracted features of voltage and current signals and dissimilarity of wavelet component coefficients are calculated, which is used to train the NN and determine HIFs from the switching operations. The SD and mean are the features considered, with 20 neurons in the first hidden layer and 10 in the second hidden layer. A total of 600 events were taken, of which 500 are used for training and 100 for testing. The combination of the NN with DWT gives a good performance. Bansal and Pillai (2007) explained that FFT is the feature extraction technique used for 320 events taken. The magnitude of the third and fifth current harmonics is used for feature vector LVQ network classifiers for HIF detection. The output layer of LVQ contains two linear neurons. In comparison with the feed-forward network with backpropagation, LVQ gives a quicker response. LVQ gives the best results in random and selected subclasses if the subclass chosen is 10.

ANN and DWT are used for HIF detection (Vahidi et al., 2010). Distorted waveforms similar to fault current waveforms are generated, and DWT is used to denoise the signal and obtain signals with a high SNR. Sym8 wavelet function is used for detection, which gives 99% accuracy. The Levenberg–Marquardt algorithm is used for training the network, with 8-5-3-1 Baqui et al. (2011). The method is robust; it uses modified HIF models and discriminates between a wide range of signals, such as HIFs. The PNN model uses a probabilistic model, such as Bayesian classifiers and a supervised learning network, a type of feedforward network (FFN) that uses exponential activation functions. The PNN structure has four layers: the input layer, pattern layer (hidden layer), summation layer, and output layer. In the network used, the initial weights and the learning process are not required (Chen et al., 2016). Fan and Yin (2019) used the convolutional neural network that overcomes the disadvantages of conventional MLPNN; for example, the spatial structure of data is not considered, and large data sets are not required for training purposes. The four layers of CNN are convolution, rectified linear unit, and max-pooling functions. Here, the classic HIF model with two anti-parallel DC sources, diodes, and variable resistors is used. Two separate conditions were studied using the proposed system: the first was tested with 5,000 data sets, of which 2,500 are HIF data sets and 2,500 are non-HIF conditions in an IEEE 34-node feeder. From the data set, 80% was taken for training and 20% for testing. The sampling rate was 15 kHz, and the input data contained 300 samples. The accuracy of the CNN obtained was 99.52% compared with the traditional MLPNN of 91.13%. The second case was performed with fewer data sets (<300) in the IEEE 13-node system with 50% training and 50% testing data. The accuracy obtained was 95.06% compared to conventional neural networks, with 74.69%. ELM is extracted from neural networks that improve feed-forward neural networks’ efficiency. ELM is a single-layer neural network, which is hidden and does not need be tuned. In ELM, input weights and biases of hidden layers are selected randomly, and the output weights are chosen analytically (Mishra et al., 2017). FFT extracts the third and fifth harmonics current and voltage magnitudes and trains the ELM (Reddy et al., 2013). The number of events is 320, and 20 neurons with unipolar sigmoidal activation functions are used for training the algorithm. Here, ELM is used for fault classification and section identification. The scheme is faster with less human intervention and gives accurate classification and less training time compared to neural networks such as LVQ and MLP. DWT is the feature extraction technique that studies the cross-countryside HIF detection applied to the transmission and distribution systems (AsghariGovar et al., 2019). Three-phase current signals are extracted at both ends of the line for fault detection and identification. The Emanual arc model is used as the HIF model, and the IEEE 13-node system is tested. The hidden layer contains 20 nodes, with the number of inputs and outputs as one. Faulty conditions and other power system disturbances are classified by ELM. The novel protection algorithm is independent of interception angle, power swing, fault location, power system topology, and noise.

A GA is an intelligent technology that detects faults (Kim et al., 1990). The behavior of the HIF is affected by many environmental parameters, and hence, a parameter-based generic technique method can be used. Detection of the HIF in a distribution system using a real coded genetic algorithm (RCGA) to analyze the tracking harmonics and current phase angles of the fault current signals was proposed by Zamanan et al. (2007). A fitness function is used by the GA that differentiates the performance between different strings. The scheme gives accurate results in differentiating harmonics and current angles of HIF. For GA, the simulation time is high, and as the processes are random, this cannot be used for fault location. GA cannot be used for online analysis because there is a possibility of inaccurate results.

Two methods for HIF detection have been compared (Sedighi et al., 2005b): the first method is the GA and Bayes classifier, where GA is used for feature vector reduction and Bayes for classification. The second method uses PCA and a NN for HIF identification. PCA and wavelet transform are used for feature extraction, for which GA gives the best results compared to the NN. The mother wavelet used is rbior3, with a sampling frequency of 24.67 kHz. It discriminates HIFs from non-HIFs, such as isolator leakage current, capacitor switching, and load switching. The tests were conducted 8.20 and 8.44 km from the source at different types of surfaces. The overall success rate of the Bayes classifier is 98.33%.

SVM is a significant classifier commonly used in issues related to different power systems and environmental fields that find application in regression analysis, prediction, and classification. SVM is a non-linear kernel-based function that maps data from one space region to another (Veerasamy et al., 2018). Different types of kernel functions are the linear kernel, polynomial kernel, radial basis kernel, and sigmoid kernel. SVM provides a distinctive training algorithm to optimize the boundaries between different groups. Optimal parameter selection is highly important to obtain successful classification outcomes. The classifier of the support vector is mainly a binary linear classifier. Theoretically, SVM was derived from the principle of statistical learning. The linear classification algorithm for the SVM applies the training set to find the segregated hyperplane. The SVM algorithm calculates the number of support vectors for SVM (no training is required), which makes SVM a better algorithm for classification than ANNs (Gururajapathy et al., 2017). The elements of the training sets that characterize the dividing hyperplane are support vectors. Even for a large data set, SVM is quick in classification with fewer heuristics. Gashteroodkhani et al. (2019) proposed a method that uses TT transform and ST for feature extraction from the transient voltage signal measured at one end. The classification is conducted using SVM optimized by a particular swan optimization and is tested over the headline and underground cable. Bewley’s lattice diagram identifies the fault location. The total number of cases was 2,376, of which 70% were used for training and 30% for testing the SVM. Compared with ST, the TT transform achieves 99.8% accuracy. Sahoo and Baran’s method (2014) uses the DWT technique for feature extraction, which considers the SD and is tested in the radial distribution feeder at 138 kV. Maximum value and energy are decomposed into two parts. Mary’s model for HIFs is used. Various classifiers, such as fuzzy, Bayes ANFIS, SVM (Kernel trick function is used), and MLPNN, are compared, among which ANFIS and SVM give the best results. A data-driven technique includes PCA, Fisher discriminant analysis, and binary and multi-class SVM for HIF detection. Compared with PCA, FDA can classify and locate HIFs successfully (Sarwar et al., 2020). PCA utilizes Hotelling’s T² statistics for HIF discrimination. The IEEE 13-node system is used for testing. Diode-resistance source is the HIF model used, with opposite polarity. Vp = 1 kV and Vn = 0.5 kV with ±10%. Rp and Rn range from 1,000 to 1,500 Ω, with random variation. In the proposed method, multiclass-SVM gives the best results. MM is the feature extraction (Sarlak and Shahrtash, 2013) and was tested on the Palash feeder in the Southwestern Tehran distribution network. An HIF indicator is installed in various poles that detect HIFs at various locations. The dependability and security of the proposed system are best above the 20 db SNR. To evaluate the proposed method, a DST, DTT is used along with PNN, as well as DWT along with SVM. The proposed algorithm has 96.9% security and 97.2% dependability.

ANFIS includes the benefits of ANN and fuzzy logic principles in a single platform. ANFIS trains the fuzzy inference system to create fuzzy rules for IF-THEN and evaluate membership features for input and output variables. A NN trains the membership functions to minimize the error in fault classification. Tawafan et al. (2012) and Abdel Aziz et al. (2011) used the FFT for feature extraction, which was then tested in a radial 13.8-kV distribution feeder. Third harmonics, the magnitude of current, and phase of current are extracted, of which the first parameter is essential for HIF detection and the latter for HIF classification. It is noted that the HIF current is < 0.05 for NCs and >0.9 for HIF conditions. The main disadvantage is determining the global minimum using the membership function and enhancing feature extraction using the algorithm. The SD values drawn from the DWT are used to train fuzzy, Bayes, MLPNN, ANFIS, and SVM and compare the performances in Veerasamy et al. (2019). DWT with ANFIS and SVM classifiers discriminates between HIFs, SLGF, LLF, and DLGF, giving superior results. Apart from this, certain performance indices, such as absolute error, root mean square error, kappa statistic, success rate, and discrimination rate, were compared. In Veerasamy et al. (2018), DWT extracts SD features from a three-phase current and is used to train ANFIS. For fault analysis, a sampling rate of 20 kHz is considered. The mother wavelet used is Daubichies’s wavelet (Db9). A radial distribution system is used for testing in MATLAB/SIMULINK. The fuzzy logic system discrimination rate is 66.6%, but ANFIS gives 100%. When discrimination rates of the fuzzy logic system and ANFIS are compared, ANFIS proves to be superior by 33.3%. The fundamental component (e.g., magnitude and phase) of current is used (Aziz et al., 2012; Abdel Aziz et al., 2012) for HIF detection, classification (LG, LL, LLG, LLLG, and HIF), and location. The fundamental component, such as third harmonics, plays an essential role in HIF detection and location. The ANFIS classifier unit gives correct output for faulty and non-faulty conditions. In fuzzy logic, the concept of possibility is used rather than the concept of probability. Jota and Jota (1999) explained the application of neo-fuzzy neurons trained to identify the SD responses where FFT is used for feature vector extraction. After the training, the neuron set becomes a decision core of the supervisory system. The method identifies the HIF in a real-time feeder, which works better for not very close fault instants. The correction index is 100% for very close faults and 81% for faults at a distance of less than 10 m. The field chosen is Caratinga in Brazil for data analysis. Silva et al. (1995) explained the HIF testing in both real-time feeders and simulation. The real-time feeder chosen was Caratinga in Brazil. The method is based on the traveling wave technique. Detection with fuzzy rules in normal and faulty cases was carried out, and faults within 10 m were identified. The SD of noise was identified to input the fuzzy. The method is focused on passive faults, which effectively work with harmonic frequencies. The feature extraction technique DWT (Silva et al., 2020) with neuro-fuzzy classifier is tested in the IEEE distribution test 13-bus feeder and the IEEE distribution test 34-bus feeder, which incorporates an evolving fuzzy neural network (EuFNN) that has an “adapt itself” ability and is excellent in terms of accuracy and robustness. Wavelet packet-based feature extraction is used with a neuro-fuzzy classifier. Different families of wavelet packet transform, namely, Haar, Symlet, Daubechies, Biorthogonal, and Coiflet, were used to extract features from a one-cycle time window of current signals. LVQ, MLP, SVM, and evolving fuzzy neural networks are compared with RMS and entropy values coefficients. Apart from other advantages, EuFNN gives a membership function for the possibility of fault occurrence. The method will not identify the fault location. DWT with a fuzzy interference system and Fuzzy ARTMAP neural network combination based on Dempster–Shafer evidence theory is tested in the IEEE-123 distribution feeder (Tonelli-Neto et al., 2017). The classification results show that the system is robust, efficient, and reliable. The Emanuel arc model is used for HIF modeling, where V_p ranges from 500 to 2,000 V and V_n from 2,000 to 2,500 V. Any new type of fault can be included in the classifier with ease. The method gives a classification efficiency of 97.69%. FFT extracts features from the signal, and fuzzy logic classifies HIF and non-HIF events (Suliman and Ghazal, 2019). The detection is performed by analyzing the third and fifth harmonics of magnitude and phase angle. The method is tested in three phase-4 wires of a 400-V radial distribution feeder in a downed conductor and wet sand. The third and fifth harmonics are extracted to train the classifier. The classifier is trained with real-time data from the practical test performed in the laboratory. By using a neuro-fuzzy interference system, tuning of the algorithm is performed.

DT represents the learned function in DT learning, which is a technique for approximating discrete-valued target functions. One of the possible values of this attribute corresponds to each branch that descends from the node. The DT-based method uses phase current (in RMS) and second, third, and fifth harmonic magnitudes to detect HIF (Samantaray, 2012). Shahrtash and Sarlak (2006) used pattern recognition with the DT algorithm, which has efficient training time and gives excellent results in even, odd, and between harmonics up to 400 Hz. Two cases are considered for HIFs: the first is when a broken conductor touches the ground and the other is when an energized conductor touches another object. The specified tree is constructed using the J48 algorithm in WEKA software. The classification factor is considered based on entropy that gives the variations in the data set. This study concludes that a 2-kHz sampling frequency and 30-cycle time interval using a small energy DT can give accurate results. EKF is used for feature extraction and ensemble DT (RF) for fault classification, compared with the DT algorithm (Samantaray, 2012). The method is tested in radial and mesh networks. The reliability and accuracy of RF is compared and is best in the two cycle window. The method is also tested in one- and three-cycle windows. For data with SNR 20, the reliability is more than 99%. The DT algorithm (Sheng and Rovnyak, 2004) can differentiate between the HIFs from normal switching, such as capacitor switching and transformer inrush currents. FFT is used for feature extraction and is tested in a radial distribution feeder. DT discriminates the extracted features for classification, which, compared with other pattern recognition tools, gives better results. Current signals of each phase are sampled at 1,920 Hz, with a total number of cases of 5,700. The Emmanuel arc model is used, and the system is simulated in EMTP.

Several other HIF detection schemes, aside from these methods, play a crucial role in HIF detection. Using signal processing and pattern recognition techniques in the device relaying architecture with expert systems, Don Russell (1990) explained low current faults. Mamishev et al. (1996) suggested using fractal geometry to analyze the chaotic properties of high-impedance defects and RMS current values are used to classify the behavior of the temporal system, resulting in a relatively short time series usable for study. Sharaf et al. (1996) used a relay-based mechanism for HIF detection. The scheme utilizes the ripple frequencies and sub- and super-harmonics usually associated with the HIF phenomenon. Eldin et al. (2007) introduced two methods for detecting HIFs in extra-high-voltage transmission lines. Both approaches investigated the origins and dynamics of HIF-related arcing. The former employs DWT analysis, whereas the latter employs the coupling capacitor voltage transformer’s high-frequency tap. The effects of fault location, interception angle, fault type, switch-off, and operations are analyzed. Both approaches are accurate and require less time.