SF6 Decomposed Component Analysis for Partial Discharge Diagnosis in GIS: A Review

This paper compiles, summarizes, and deliberates over one hundred important works on the different approaches and advances in the surveillance and diagnosis of the internal status of SF6 gas-insulated equipment (GIE), particularly partial discharge (PD) diagnosis in gas-insulated switchgear (GIS), and the proposed diagnosis techniques used. This review focused on four research aspects on PD diagnosis related to the developments in PD detection techniques, PD sources identification, and PD severity evaluation. Besides, the effect of various factors such as gas pressure, applied voltage, and impurities on the deterioration of insulation gas and its influence on the diagnosis process has been reviewed. Currently, some reviews on PD diagnosis in SF6-insulated switchgear have been presented and analysed; however, to date, most of them tend to focus on various PD detection techniques in GIS, while others are not extensive and comprehensive reviews. Unlike the available review publications, this paper highlighted various aspects of PD diagnosis in GIS and created a base for further development. The research trend in this field is expected to be directed toward a comprehensive assessment. This review provided a position of the current PD diagnosis in GIS studies and developments that can be a guideline for researchers for further research on the topic’s actual impact in the field.


I. INTRODUCTION
SF6 gas is widely used as an electrical insulator in power systems due to its high dielectric properties and excellent arc quenching capability [1]. SF6 has passed many qualification tests required for insulating gas in high voltage (HV) systems. It has become the most common insulation gas in modern high voltage, extra-high voltage, and ultrahigh voltage power system equipment [2]. These equipment include gas-insulated switchgear (GIS), gas-insulated transformer (GIT), and gasinsulated transmission line (GIL). Recently, the SF6 market has expanded, with the electrical sector accounting for around 80% of total SF6 demand [3][4] [5]. Due to its wide usage in the power system, monitoring the performance of SF6insulated equipment is critical for a safe and reliable power system operation.
When subjected to electrical discharges, such as due to an electrical fault, portions of SF6 gas dissociate or decompose into by-products, which in turn affect its overall insulation properties. Apart from electrical discharges, SF6 decomposition may also be caused by thermal stress [6] [7]. Many researchers have studied the high current discharges, such as the arc and spark discharges, as well as the small current discharge or the partial discharge (PD), and the impact of the discharges on the insulation strength of the gas-insulated equipment (GIE) [8] [9][10] [11]. Various techniques were employed to detect the discharges, especially the partial discharge, including the acoustic, chemical, and ultrahigh frequency (UHF) techniques. There have also been significant attempts to apply different AI approaches to recognise the PD sources and PD types, such as support vector machines (SVM) and wavelet transformations [12][13] [14]. It is known that the PD pattern is a significant indicator to identify the type of PD source within a GIS [15]. In addition, it is also desired to be able to evaluate the degree of PD severity because this critical information can help in overall equipment condition monitoring. Previous research reported that the number of decomposed by-product gases and their concentration could reflect the severity of PD activities in GIS [16] [17]. However, a specific guide for the PD severity evaluation is yet to be established [18]. In short, it is desired to have regular PD assessment of a GIE and early detection of possible insulation degradation may prevent catastrophic problems from occurring.
Recently, several techniques for monitoring and diagnosing GIS were proposed and developed. An illustration for these techniques, their working principle, and drawbacks are required. There are some review publications regarding these techniques available in the literature, which included trends and state of the art in particular aspects [19][20] [21]. However, it does not comprehensively review developments in PD detection techniques, PD severity evaluation, and factors that cause SF6 decomposition.
The motivation of this paper was to review recent monitoring and diagnosing development and trends toward a comprehensive evaluation of the SF6 GIE, particularly GIS. By focusing on publications from the last decade, this review highlighted the causes of SF6 deterioration, revealed PD detection methods, clarified various techniques used for PD and insulation defects identification, and provided a theoretical basis for current severity evaluation approaches. This review also presented a taxonomy for multiple strategies adopted in the literature and a starting point for further research on this topic, on top of identifying relevant gaps.

II. INSULATION CHARACTERISTICS AND DECOMPOSITION OF SF6
SF6 is a highly electronegative gas that attracts free electrons, resulting in a dielectric strength three times that of air at atmospheric pressure. It has good arc quenching properties and excellent insulation and heat transfer characteristics, making it an appropriate insulator in GIE. Pure SF6 is an inert gas that is colourless, odourless, nontoxic, and noncombustible. Since it does not decompose at temperatures below 500 °C, pure SF6 is heat-stable and has good thermal conductivity [22][23] [7] [21].
It is known that SF6 gas decomposes when subjected to an electrical discharge where the extent of the decomposition is related to the intensity or energy of the discharge [24] [25]. Previous investigations of SF6 decomposition mechanisms under PD partial discharge were reported in [26] [27]. To demonstrate the decomposition mechanisms under a negative corona, the authors proposed the case of a point-plane electrode configuration located in an SF6 gas chamber. Three main regions were identified as the glow region, the ion drift region, and the main gas volume region as shown in Figure 1. In the glow region, SF6 gas ionisation and dissociation are mainly caused by electron collision with SF6 molecules. The ionisation process produces several low fluorine sulfides, namely, SF5, SF4, SF3, SF2, and SF. In addition, oxygen atoms (O) and hydroxides (OH) may also be produced by the dissociation of oxygen molecules (O2) and water vapor (H2O), respectively. Several additional components may be generated by the interaction of the above substances in the glow region. These include long-lived gases such as sulphur dioxide (SO2), thionyl fluoride (SOF2), sulphuryl fluoride (SO2F2), and disulfur decafluoride (S2F10). Nevertheless, SF6 molecules may also quickly reform due to self-healing process involving fluorine (F) atoms and low fluorine sulfides. Equations (1) to (12) illustrate the key reactions that occur in the glow area [26][27] [28]. e + SF6 → SFx + (6 -X)F + e, X < 6 (1 The ion drift region is located between the glow and the anode plane. In this region, negative ions are transported and then discharged at the anode. The volume encircling the pointplane area is the primary gas volume region. In this region, the chemistry is characterised by a slow gas-phase or surface reaction. As previously mentioned, in the glow region, the presence of SF6 decomposition by-products and other impurities, such as O2 or H2O molecules, leads to additional components to be produced. The additional components are produced due to further reactions in the main gas volume as described in (13) to (16 Figure 2 summarises the formation of SF6 decomposition by-products. More investigations on the decomposition mechanism and how to obtain decomposition characteristics under PD caused by various insulation defects are needed.
From the figure, the high-energy electron stream collides with SF6 molecules under the influence of PD, resulting in gas decomposition into low fluorine sulfides. This collision splits S-F bonds, resulting in active F atoms and unstable SFx primary products [29]. Some reactions occur between these products and traces of H2O and O2, leading to multiple decomposition by-products.
In another research [30], a study was conducted to examine the behaviour of AC corona discharge on the decomposition mechanism. The author applied 50 Hz AC corona to a mixture of SF6 gas and H2O. Their findings showed that 50 Hz AC corona acted similarly to negative DC discharge.

III. MECHANISMS OF SF6 DECOMPOSITION
Despite its excellent properties, SF6 gas may decompose into various by-products as a result of an electrical discharge, as previously noted. Apart from electrical discharges, SF6 gas may also decompose as a consequence of thermal overheat, also known as partial overheating fault (POF), and x-ray irradiation [6] [7]. The various SF6 decomposition mechanisms are summarised in Figure 3. In the order of reducing energy, the electrical discharges can be further classified into arc, spark, and partial discharges [31] [1].

A. ARC DISCHARGES
Arc discharges are characterised by a current of a few thousands amperes, a discharge period from a few dozens to a few hundreds of milliseconds, an energy of about 10 5 -10 7 J, and a temperature of up to 20,000 K [22] [32]. Power arc is generated during current interruption in circuit breakers, short circuits inside gas chambers, and disconnectors operation [33]. The SF6 concentration in a GIS usually declines after an arc discharge. The recovery of SF6 after its decomposition is especially difficult due to the presence of impurities, such as water vapour and oxygen, as well as metal and carbon originating from electrodes and erosion [34]. Moreover, some newly generated decomposition by-products maybe corrosive, resulting in the corrosion of solid insulation materials and in a reduction of the overall GIS dielectric strength. Other effects of arc discharge on the SF6 decomposition mechanism in the presence of various impurities can be found in [34]. After an arc discharge, as the arc cools, sulfur and fluorine atoms may recombine to form SF6 molecules. However, when the arc temperature drops from 12,000 K to 300 K, other by-products may also be generated [35] [36]. In a high-energy discharge, the SF6 decomposition by-product is mainly SOF2, as well as a small amount of SO2F2 [31] [28]. However, decomposition process is affected by impurities and materials included in the equipment.
Limited studies have carried out to investigate the influence of some traces in SF6 decomposition under arc discharge. In [37], the effects of trace H2O and O2 impurities and polytetrafluoroethylene (PTFE) vapour on SF6 by-products were studied. Results obtained have found that the main products after arc quenching were CF4 under PTFE, SOF2, SO2 and SO2F2 when O2 is involved, and SOF2, SO2 and HF when H2O is involved. These findings emphasis the influence of various factors involved and its relevance to the decomposition process under arc discharge. Therefore, further studies to confirm the obtained results under various arc energies are needed.

B. Spark Discharges
Unlike an arc discharge, a spark discharge has a lower energy level, which is approximately 10 -1 -10 2 J, and a longer discharge period. It is noted that the discharge time affects the gas decomposition by-product's volume [22]. The spark's energy magnitude is three orders that of the corona [38] [39]. The sequence of electron avalanches that started at the cathode is crucial in spark discharge to create the gas breakdown [39]. Spark discharge is originated from higher field strength than that the dielectric breakdown field strength of GIE, which is caused due to braking actions of disconnectors, fast transient overvoltage caused by ground faults, insulation defects inducing high local electric field strength, and continuous rising of equipment voltage [17]. Following a spark discharge, the organic insulation becomes carbonised by the high temperature spark, resulting in a decline in the overall dielectric strength and may pose a threat to the safety and stability of the power system operation, in addition in extreme cases, the metal electrodes may be eradicated [38] [40].
Similar to other electrical discharges, a spark discharge may produce several SF6 decomposition by-product gases. In a decreasing order of the quantity produced in a spark discharge, the by-product gases are SOF2, SOF4, SiF4, SO2F2, and SO2. Several other studies had also found S2OF10, S2O2F10, and S2O3F6 as by-products of a spark discharge in SF6, and S2F10 is regarded as the difference between spark discharge and PD or arc discharge despite its low content [41] [31] . Meanwhile, decomposition process and generation of decomposition products under spark is affected by several factors, such as impurities traces of H2O and O2, gas pressure, solid insulating and electrode material, electrode.
It is known that the gas pressure affects the accumulation of electron energy in a spark channel, causing electron intensity to be diverted into SF6 [40], however, in terms of the effect of gas pressure on the by-products produced by a spark discharge, the study found no such effect can be detected. Another investigation on the impact of H2O on SF6 dissociation under spark discharge was also performed [17]. Their results revealed that H2O had a noticeable effect on the production rate of the decomposition by-products, namely, SOF2, SO2F2, SOF4, SO2, and CF4. On the other hand, O2 greatly influenced the growth of SO2F2 and SOF4 but not SOF2 [38]. In general, previous studies have shown that impurities presence has a noticeable influence on the decomposition process, which determines the key feature parameters to recognize spark discharge in gas compartment. However, a diverse results of decomposition products and their content have obtained, thus, further evaluation of the relation between various impurities content and generated products is essential for spark discharge diagnoses in GIS.

C. SF6 Decomposition by Thermal Effect
Apart from the electrical discharge, the presence of heat can also cause the decomposition of SF6 gas. For example, thermal decomposition of SF6 occurs when a high current flows through an oxidised layer's contacts, which have high resistance and poor electrical conduction and hence overheating. Overheating can also be caused by other factors, such as electrical faults or short circuits caused by damaged insulation or current overload. Similar to the electrical discharge, thermally induced decomposition process produces several low fluorine sulfides, namely, SF5, SF4, SF3, SF2, and SF [5]. It is noted that thermal decomposition decomposes not only SF6 but also the organic insulating materials such as the epoxy resin. As a result, H2O molecules could be released and react with SF6, this reaction develops the thermal decomposition of SF6. In addition, the organic insulating materials attract F atoms from SF6, resulting in low fluorine by-products [42][43] [44]. In a decreasing order of the quantity produced in the thermally induced reaction, or partial overheating fault (POF), are SO2, SOF2, SOF4, SO2F2, and H2S. SO2 and SOF2 account for 90% of SF6 decomposition under the POF. Also, SO2 is usually the first by-product produced, and its content increases with temperature [41] [45]. Nevertheless, the generated products and their content vary according to the temperature at the thermal fault location in GIS.
Previous researches have investigated the relationship between thermal fault temperature and the SF6 decomposition process. An experimental study detected decomposition products under POF at a temperature lower than 400 °C [44]. Results showed that the most extensive products were SOF2 and SO2. The concentration of SOF4 and SO2F2 was minimal, and H2S was detected at temperatures above 340 °C. Furthermore, due to the above effect of organic insulating materials under overheating, decomposition of solid insulating materials has also been investigated under varied temperatures [42] [43]. Results had revealed that at 250°C, only H2, CO, CO2, and CH4 can be detected. Whereas by rising temperature to 300°C, SOF2, SO2, H2S, COS, and CF4 have been detected. In brief, the above studies have shown that SF6 thermal decomposition is closely related to the thermal fault temperature. However, the decomposition should consider not only direct thermal fault impact on SF6, but also the presence of organic material, especially at various temperatures to propose consistent characteristics parameters for thermal fault in GIS.

D. SF6 Decomposition by X-ray Irradiation
SF6 gas can also be decomposed by x-ray irradiation. Similar to electrical discharge and thermal induced decomposition, SF6 decomposes when exposed to x-ray irradiation. In certain applications, x-ray units are used in the vicinity of SF6insulated equipment for the purpose of equipment defect detection using digital imaging technology [46][47] [48]. Xray has been introduced as a defect detection technique due to its ability to penetrate and visually provide results for the internal status of GIS. The decomposition by-products produced by the exposure of x-ray are mainly similar to those produced under electrical discharges such as S2F10 and SOF2 [47].
The relative generation rates of S2F10 and SOF2 are approximately the same as in PD condition, however, S2F10 generation rates are nearly a factor of ten less than that noticed for SOF2 [47].

E. SF6 Decomposition by Partial Discharge
According to IEC 60270, partial discharge (PD) is defined as a localized electrical discharge that only partially bridges the insulation between conductors and which may or may not occur adjacent to a conductor. Insulating materials in high voltage electrical equipment can deteriorate chemically and physically as a result of PD activity. Being a type of partial discharge, the term corona discharge is used when the discharge occurs on the equipment's external surface. However, in GIS discharge monitoring, the terms PD and corona discharges are used synonymously [49].
In a GIS, PD is generated due to commonly found defects such as sharp protrusion tips on the current-carrying conductor or the metallic enclosure [50] [24]. When a sharp protrusion is present, its surrounding electric field may be distorted, leading to a strong local electric field. Hence local discharges can easily occur around such a sharp protrusion. A stable PD is initiated when the GIS is energised at standard operating voltage. A complete breakdown of the GIS may eventually occur after a certain period of continuous PD due to SF6 gas degradation, or when the GIS is under an overvoltage. Consequently, due to the PD activity, the equipment's withstand voltage decreases considerably [51]. Untreated defects cause insulation deterioration to aggravate, and PD activity gradually leads to a severe spark fault. Apart from a sharp protrusion, other defects or sources of PD in GIS include free conducting particles, floating components, voids, and contaminations on the solid dielectric and on spacers used to support high voltage conductors [52] [53]. The free-moving non conducting particles may also become dangerous when mechanical vibration occurs due to the electrostatic force, and hence inducing PD by their movement [9].
In GIS, PD is considered an indicator of dielectric's degradation and the initial cause of the deterioration rate's acceleration. The types and concentrations of decomposition components are strongly affected by defect severity, applied voltage, discharge duration, and gas pressure [54]. Despite its low energy compared to arc and spark discharges, and its low discharge current in the order of microamps, PD represent 85% of causes of GIS catastrophic failures. When the PD occurs over a long period, many decomposition products are generated [22] [20]. For example, a corona discharge in SF6 gas produces SO2, CF4, SOF4, SO2F2, HF, SOF2, and CS2 [39]. Numerous previous studies were conducted to understand the decomposed components for better GIS monitoring. Nevertheless, PD diagnostic in GIS by studying SF6 decomposition components is a research topic that still requires further investigation.

F. Summary of Main Decomposition Products Under Electrical Discharges
Various electrical discharges could cause an extensive dissociation of SF6 into its lower sulfur fluorides. Further reactions of these sulfur fluorides with traces of residual moisture or oxygen lead to the generation of further byproducts. The type of by-products and their generation rate vary according to the electrical discharge type. Table I summarises the main SF6 decomposition by-products under main discharges based on the discharge energy. It is noted that each component's production rate varies according to the decomposition mechanism, making it a helpful tool to identify the internal deterioration type.

IV. PD DETECTION METHODS IN GIS
The GIS integrity and reliability are crucial to the overall safety and continuity of a power grid. As previously noted, partial discharges occur in a GIS even at the normal operating voltage, and these discharges may eventually lead to the GIS failure. Hence, early detection and assessment of PD in a GIS is paramount. There are various methods available to detect the PD occurrence.
When a PD occur, electromagnetic waves, high-frequency switching pulses, light, sound, and decomposition byproducts, may all be accompanied, and serve as measurable quantities for PD detection and hence, insulation diagnostic [19] [55]. The available methods used to detect PD are categorized as either conventional, that is based on current or charge measurement as described in IEC60270 standard, or nonconventional, that is based on other quantities or physical phenomena such as electromagnetic waves, light, sound, and decomposition by-products [56][52][50] [57].

A. Electrical Methods
The electrical PD detection is based on the detection of highfrequency current pulses, electromagnetic waves and apparent power loss that accompany the occurrence of PD. The pulse currents can be measured using a discharge free high voltage coupling capacitor in series with a low voltage arm (high pass filter), high-frequency current transformer (HFCT), or a Rogowski coil. The pulse current detection method (IEC60270) is the only technique that can measure PD quantitatively in pico Coulombs (pC). However, due to its low resistance to electromagnetic interference, it is inappropriate for online measurement [58][50] [59].
In GIS, the electromagnetic wave generated by PD activity can induce a transient voltage on the cylindrical wall of the GIS, also known as transient earth voltage (TEV), which can then be measured using a capacitive probe or divider [60]. The TEV is formed on the metal surface in GIS due to highfrequency electromagnetic waves spreading across the insulator and flange junctions, representing discontinuities in GIS and allowing high-frequency PD to propagate to the outer metal. However, electromagnetic interference and peripheral noise could affect the PD detection using this method.
The HFCT is also employed for PD detection by clamping around cables and GIS case ground [21] [19]. As previously noted, PD detection using HFCT is based on the detection of PD pulses-induced currents. These currents mainly flow through the inner compartment part of GIS and the primary conductor's outer part. PD signals far from the PD source could also be measured using HFCT due to the lower attenuation of induced currents [61].
The electromagnetic waves generated by a PD activity can be conveniently measured using suitable antennas such as flat or horn shaped antennas. In a gas insulated switchgear (GIS), the antenna is in the form of a disc and is located at suitable places along the GIS cylinder, such as at the interfaces. These are also known as UHF sensors or disc couplers. The UHF detection range is mainly between 0.3 and 3.0 GHz [62]. The first stage in UHF PD measurement is to obtain electromagnetic signals for further signal processing. Hence, the accuracy and sensitivity of measures will be significantly influenced by UHF sensor characteristics and system configuration. Figure 4 shows a block diagram of UHF PD measurement [57].
The structure of UHF sensors varies according to the antenna applied for signals detection. In GIS, UHF antenna could be installed internally to the inner surface of the compartment, or externally which has been developed and utilised for PD monitoring. It includes windowed and barrier couplers installed at different locations [57] [62]. With the external installation, the integrity of the insulation system is less affected.
The UHF method has several advantages, including high sensitivity and good anti-interference capability [58]. The maximum attenuation of UHF signals in GIS is five pC. However, the UHF approach is ineffective for detecting mechanical vibrations [21]. It also has problem with discharge calibration, and low-energy discharges are challenging to notice [32][50][59] [63].
Another type of electrical method for PD detection is through the use of apparent power loss or loss tangent measurement. However, this measurement only gives the bulk characteristics of PD activities in the form of apparent power loss measured using Schering bridge. It is useful to determine conditions of equipment such as motors and generators.
Among the electrical detection methods, pulse current is the only one that could provide quantitative determinations [60]. In short, the electric methods have the limitations of interference, sensitivity, and its ability for online detection.

B. Acoustic Methods
The resultant mechanical vibration and ultrasonic waves with long wavelengths, strong directionality, and concentrated power produced by PD and particle movement can be detected using acoustic sensors. GIS uses it to identify insulation defects by extracting parameters from detected acoustic signals by an appropriate acoustic sensor [64].
The most widely used acoustic sensor is the piezoelectric transducer. It includes a polarised material that serves as the active element. Mechanical vibrations will be converted to electrical signals by this sensor and vice versa [65]. Acoustic sensors are immune to electromagnetic interference, but lowenergy discharges are not detected at the preliminary stage due to acoustic signal attenuation [32]. It is also easy to install on the equipment's surface [21]. However, noise can significantly affect ultrasonic detection, leading to inaccurate detection and identification of PDs [59] [50]. The precision of the acoustic and electrical methods heavily relies on an adequate signal-tonoise ratio [63].

C. Optical Methods
PD activity also produces light radiation from current flow, in addition to electromagnetic radiation, and acoustic waves [66]. Two types of optical PD detection systems are becoming increasingly popular because of their high sensitivity and immunity to electromagnetic interference (EMI) [67]. In the first type, the optical signals produced in GIS are directly detected. Sensors of this type have been installed in the GIS. As a result, this method is effective when the discharge source generates light. However, this method is invasive. The second type is the detection of acoustic emissions based on optical sensors. In this method, the sensitivity of the acoustic emissions is improved using optical sensing to overcome the sensitivity limitation of traditional acoustic emission detection.
A vacuum photomultiplier tube (PMT) and a high-speed intensified charge-coupled device (ICCD) were used for optical PD measurement. However, their application was limited to laboratory studies and was hardly applicable to practical SF6 insulated equipment. This drawback can be improved by employing a single-photon level photosensitive technique. On a millimetre scale, the silicon photomultiplier (SiPM) is a single-photon sensitive sensor made up of thousands of single-photon avalanche diodes (SPAD). Its competitive features include a low bias voltage, high quantum efficiency, and compact device size, making it a viable replacement for other photoelectric sensors such as PMTs [68] [69].
A double spectral SiPM sensor technique to diagnose PD in GIS was reported [68]. Figure 5 shows the schematic diagram of the synchronous PD measurement system. The results demonstrated that the SiPM sensor had the same sensitivity to PMT detection and HFCT. However, SiPM sensors required internal installation, and further confirmation is needed in real GIS. Therefore, an optical fibre sensor was proposed by [67] to detect acoustic emissions. In an experiment using a 126 kV GIS, the performance of an optical-fibre ultrasonic sensor and a conventional lead zirconate titanate (PZT) sensor was compared.
The proposed optical fibre sensor method revealed that the number of detected PD was more significant than PZT and had a higher signal amplitude. However, these sensors have yet to be applied in existing GIS. The optical measuring method was not hampered by live communication. The optical sensor's numerical aperture limiting angle still exhibits sensitivity and accuracy in locating the discharge source [21]. This procedure is still in its early stages of development [70] [19].

D. Chemical Methods
This chemical diagnostic technique which is known as decomposed component analysis (DCA) or chemical byproduct analysis, has attracted the attention of researchers.
This method is based on the analysis of SF6 decomposition products generated by the PD activity. The presence of PD in GIS may decompose SF6, resulting in various products with different concentrations. The type and concentration of the SF6 decomposition components in GIS could provide sufficient information on the degree and type of internal degradation before breakdown. Researchers discovered an approximately linear relationship between PD energy and the amount of SF6 decomposition products. Product types and quantities are related to H2O and O content, discharge quality, electrode materials, and so on [21]. Examples of gaseous or solid products include SOF4, SO2F2, SOF2, S2F2, SF2, S2F10, CF4, CO2, SF4, S2OF10, H2S, HF, SO2, SiF4, WF6, and CUF2 [71].
This method has a significant potential for PD early warning with high sensitivity and factuality. Noise and EMI have a rare effect on chemical analysis, making them a useful diagnostic tool for detecting PD. The procedure is noninvasive and can be performed online [50] [35][63] [7].
Methods used to analyse SF6 decomposition by-products are Fourier Transform Infrared Spectroscopy (FTIR), gas chromatography (GC), detector tubes, and gas sensors [7] [71]. The accuracy of these techniques and their ability to detect decomposition products is essential for accurate PD detection using the chemical method.

1) Gas Detection Methods
The detection techniques of SF6 decomposition products provide the required data to implement the DCA method and PD diagnosis. These techniques can detect the concentration of decomposition products in parts per million (ppm) level. However; their accuracy and ability to detect the types of products are varied. The working principle, advantages, and drawbacks of the tools have been discussed.

i) Detector Tube
A detector tube is an analytical method that utilises chemical reactions in a test tube to identify the decomposition products [20]. Detector tubes are typically used for on-site detection, but large sample volumes and high cross-sensitivity defects make it difficult to detect gases correctly [64]. Certain decomposition products, such as SOF2, SO2, and HF, can only be observed when commercial detector tubes analyse SF6 gas decomposition by-products. A detector tube's accuracy can even exceed μL/L. However, temperature and humidity do not affect its stability. Besides, cross-interference exists due to positive interference when some compounds are chemically similar to the target compound [7].

ii) GC
GC is a gas analysis method that is based on column separation. The process begins by vaporising the sample mixture before a carrier gas such as nitrogen, hydrogen, or helium transports the sample through a chromatographic column. The flow of the sample with the carrier gas causes the samples' components separation. The separated components are then sent into a detector to record signals and transform them into a chromatograph. Components separation occurs at different times, which is referred to as the retention time [72]. Figure 6 shows a block diagram of the main parts of GC. The first part is the injector, through which the sample is introduced. The chromatography column is placed in the oven, while the detector provides the signal of separated components [73].
Although GC has excellent accuracy, it requires periodic maintenance; hence cannot continuously track the gas [63]. With precision, the GC system can detect most components, such as SO2, SOF2, SO2F2, and CF4, which exceeds μL/L. Nevertheless, the sample injection time is exceptionally long, and the chromatographic columns require frequent cleaning. Therefore, this approach is unsuitable for online surveillance [7].

iii) FTIR
The FTIR is the most widely used infrared spectrometer. It is based on infrared wavelength range measurements that the sample absorbs after being exposed to infrared radiation from an infrared source. The presence of reference spectra is necessary for quantification [74] [75]. The FTIR detection technique has the advantages of being fast, could detect various component types, have high detection accuracy, resist disturbance, and identify the same sample repeatedly. Therefore, this technique is appropriate for online surveillance [7]. However, the dynamic configuration and high costs often hinder the practicality of FTIR and GC, making them more useful for offline laboratory research than on-site [63].

iv) GAS SENSOR
An operation of a gas sensor is based on the conversion of measured gas concentration to an electrical quantity. The electrical properties of sensor materials can change due to the chemical properties of sensor material after it absorbs a gas molecule to its surface [76] [77]. The advantages of this gaselectric conversion sensor are its high-speed and performance. For SF6 decomposition product detection, the gold(Au)-doped titanium dioxide (TiO2) nanotube array sensor was proposed [78]. This study observed an improved gas sensing performance and reduced operating temperature for the TiO2 nanotube array sensor. Phosphorene was also suggested to probe SF6 decomposition gas sensors as it shows a variable response to SO2, H2S, and SF6, revealing a promising material for online GIS diagnosis [79]. However, sensor poisoning occurs after prolonged use due to the gases' chemical reaction to be measured and the sensor. Thus, the sensitivity and accuracy of the gas sensor could be affected [79][80][69] [81].

v) Ultraviolet (UV) Spectroscopy
UV detection is used in the 200 -400 nm range of near UV light. The UV spectroscopy detection system consists of a fibre UV spectrometer, a deuterium lamp, transmitters, and a sample gas. Noninvasive measurements and a small scale in transformer diagnosis are among its advantages. However, UV absorption features of main by-products have not been observed, and the strict vacuum environment of detection restricts its applications. Thus, applying this technique for early on-site detection of PD is a topic worth investigating [63].

vi) Overview of Gas Detection Techniques
Electric gas sensors, FTIR spectroscopy, or GC may acquire more precise quantitative data. However, for on-site real-time monitoring, their practicability needs to be improved [63]. More research is required to develop a consistent detection system for detecting SF6's decomposition under different conditions. PD detection based on the chemical method using SF6 decomposition by-products is a relatively new field. Various types of chemical analysis equipment have roughly different detection ranges for gases and their concentrations. Hence, more investigations using a chemical approach is required. Also, the relationship between generated gases and gas concentrations of decomposition by-products and the PD energy levels at various types, quantities, sizes, and materials of the defect must be reported. Table II summarises the decomposition by-products detection techniques.

2) PD Analysis Based on Decomposed Component Analysis (DCA)
Decomposed component analysis has received attention as a promising diagnostic tool for GIE internal conditions. By qualitative and quantitative surveillance of the decomposition products, faults monitoring and the internal insulation condition can be realized, due to the close relation between generated products and their concentration with the type of insulation degradation [6]. Detecting the SF6 decomposition products in the gas chamber refers to the presence of a PD source in the equipment [9]. The relationship between decomposition products and PD conditions has been investigated using the chemical detection method. For instance, in [54], the optical properties of the decomposition products have been evaluated using FTIR under different PD conditions. It's found that the content of CF4, SOF2, and SO2F2 has increased along the discharge period. The content of SO2F2 has increased with the increment of discharge voltage.
The process of PD analysis using the DCA method starts with the data collection using decomposition gas analysis tools such as FTIR and GC, and followed by the processing and selection of feature parameters; that could better improve the interpretation of PD activity. The selected characteristics features are then applied for AI techniques to monitor and diagnose PD activity within GIE. The characteristic decomposition products such as SOF2 and SO2F2, products ratios such as SO2F2/SOF2 and sulfur to carbon ratios, and the application of the artificial intelligence (AI) techniques such as SVM and tree-based algorithms are among the current PD data analysis techniques based on DCA method. Figure 7 shows the processes included in PD analysis using chemical analysis and AI techniques. Future studies on PD analysis using DCA are still needed due to limitations that are related to feature parameters extraction, selection of analysis techniques for source classification, and severity evaluation. VOLUME XX, 2017 9  [63]. Moreover, none of these techniques can be utilised individually to eradicate the surrounding interference as the PD signals collected from each approach only include a specific facet of the information.

E. Overview of PD Detection Techniques
Recently, combined detection methods have been suggested by some researchers to improve the performance of diagnosing and monitoring systems while overcoming the limitations of a single approach [62] [82]. Consequently, effective diagnostic techniques are necessary to ensure high reliability in detecting defects that cause insulation deterioration. Figure 8 categorises the different methods for detecting PD in electrical equipment.

V. CLASSIFICATION METHODS OF PD SOURCES IN GAS INSULATED EQUIPMENT
Various AI techniques were used to intelligently classify PD defects in GIE when digital electronics and signal processing methods evolved. Significant efforts have been made to employ AI techniques such as artificial neural networks (ANN), wavelet transformation, and support vector machine (SVM) to classify PD sources automatically. PD classification in GIE have reviewed based on studies employing both nonchemical and chemical analysis.

A. Non DCA Based Classification
This classification approach uses techniques such as the PD measurements, PD phase resolution (PRPD) pattern and X-ray irradiation to identify the types of defects in GIE. X-ray digital imaging technology was applied to detect various defects: metal and flaky particles, Loosened Metal Screws, and different adsorbent cover materials [46]. The investigation concluded that the technique for visually detecting defects in GIS was effective. However, SF6 may decompose due to Xray irradiation by employing this technique, resulting in a significant reduction in SF6 insulation. Consequently, more research is required to investigate the influence of X-ray on SF6 gas before it can be applied on-site.
Moreover, a probabilistic neural network (PNN) was developed to classify PD produced by a floating electrode, polluted bushing, and free metallic particles in gas-insulated load break switches (GILBS). As for data reduction, the fuzzy C-mean method was employed [15]. Multisource PD classification was conducted by [83] using two-level logistic regression, with the feature extraction of PD signals was based on PRPD. The author revealed that a single class SVM could classify individuals based on multisource signals.
These methods have performed well; however, the accuracy of PD measurements was subjected to measurement and sensing devices ability [19][84] [85]. Methods used to measure PD have limitations due to interference and noise accompanied by PD signals. Consequently, using noise removal techniques for better pattern recognition are required. Also, feature extraction of pertinent data from raw data is needed during the preprocessing stage.
Alternatively, chemical by-products might be considered by using SF6 by-product concentration and ratios as feature characteristics to overcome the limitations of PD measured quantities.

B. DCA Based Classification
The concentration and concentration ratio of SF6 decomposition products were used as feature parameters to classify insulation defects in GIE. A decision tree algorithm with concentration ratios as a classification parameter was suggested to detect GIS defects. A protrusion, free conducting particles, contamination, and gap were used as artificial defects for PD initiation [50]. However, decision tree faces the challenge of overfitting. Another author has used PSOoptimised SVM to discriminate four defects within a GIS, namely protrusion, particle, contamination, and gap defects. The author proposed three concentration ratios as a feature parameter for classification [52]. However, SVM performance was mainly dependent on algorithm parameters that the user must specify. The SVM is also a binary classifier. For multiclass problems, the classifier should be generalised.
Meanwhile, the backpropagation neural network (BPNN) algorithm was used to diagnose DC-GIE using feature sets of SF6 decomposition quantities [16]. The authors compared the accuracy rate of identifying four types of defects (protrusion, free conducting particles, polluted insulation surface, and insulator gap) using concentration and its ratio from the suggested feature sets. Results found that the classification showed a high accuracy rate with the concentration ratio feature set. However, BPNN has a drawback related to overfitting if not generalised.
Meanwhile, SF6 decomposition under three categories of PD sources was studied [86]. In the research, the random forest algorithm was employed. The outcome demonstrated a high classification accuracy compared to eight algorithms used to classify various defects utilising product concentration as a characteristic quantity. Based on the Duval idea used in the dissolved gas analysis (DGA) in transformers, the authors in [41] and [6] applied the triangle method to identify PD, POF, and spark fault in GIE. The concentrations of four decomposition products were selected to identify the three fault types to build the triangle sides. They found that a high accuracy rate was detected in the internal fault recognition using the diagnostic method. The researchers also concluded that it is possible to provide an on-site diagnosis for power equipment.
The triangle method with the decomposition product concentration and concentration ratios was also used for discharge recognition caused by the metal protrusion, floating, and insulator defects [87]. This graphical method had demonstrated reliability and sensitivity in recognising incipient faults. However, further confirmation is needed in real GIS.
Furthermore, the feasibility of recognising protrusion, particle, contamination, and gap defects using decomposition characteristics was investigated by [28]. After analysing the decomposition products, the author concluded that the four types of defects differ significantly in decomposition amount, decomposition rate, and concentration ratio. However, this classification depended on the user's knowledge of the decomposition product characteristics at each type of defect. A diagnosing method that combines two detection techniques for PD was proposed by [59]. The author used the UHF method for online detection while chemical analysis for offline detection. The obtained data with the Dempster-Shafer (DS) evidence theory application was used to classify four types of GIS defects. The results obtained using this combined method showed an improvement in diagnosing accuracy.
There is still a need to test the method's performance in the gas chamber with simultaneous PD sources. Furthermore, more investigation of other combined detection techniques is required to compare performance with the proposed method. Table IV illustrates some of the methods used to identify defects in GIE using chemical analysis, as well as its limitations.
Previous works had been successful in detecting defects within GIS to some extent. However, applying AI techniques for PD source classification faces challenges in selecting appropriate features and a suitable classification algorithm, especially in multiple PD sources.

VI. DECOMPOSED COMPONENT ANALYSIS BASED PD SEVERITY EVALUATION
An insulation status assessment includes determining the internal deterioration type and its severity [88]. Hence, the degree of severity has attracted researchers' interest based on the essential information obtained. Studies on PD severity have been conducted across single and multiple types of defects. However; most of these studies have focused on PD severity induced by the protrusion defect, which is a common and harmful defect in GIS. The protrusion has been simulated using a needle plane and point plane electrodes. The use of characteristics parameters extracted from experimental and theoretical studies to identify the internal operation condition in GIS can also be employed to evaluate the severity of the discharge [51]. Previous studies have found that the decomposition products' quantity, concentration, and ageing time variation can reflect the severity of PD in GIE [16] [17]. However, the criterion for PD severity evaluation has not yet been unified [18].
Several approaches have been investigated to evaluate the severity of PD. For instance, as decomposition occurs, SOF2 and SO2F2 are generated [Equations (13) and (14)] due to the reaction of SF4 and SF2 with H2O and O2, respectively. Since the energy required to generate SF2 is more than the energy needed to produce SF4 to break the bonds S-F in the SF6 molecule, SO2F2 content generated at a higher PD level is more significant than SOF2. Thus, the ratio (SOF2 + SO2)/SO2F2 can be used as a severity indicator. In such a way, the generation of decomposed components will define the PD's severity in GIE [89].
The results indicated that SO2 concentration increases with PD improvement, whereas H2S became a significant product that only generates high-energy discharges. Therefore, these two products can characterise insulation deterioration severity [90][63] [29]. Besides, since the generation rate of decomposed components was associated with the severity of discharge [91], it can assess the severity of the defect. For statistical representation, the effective generation rate was considered [10]. Another study also reported using the enhanced concentration ratio (SO2F2)/(SOF2) by using the effective characteristic ratio to identify the PD energy statistically [29].
Furthermore, in another study, the changing levels of nonuniform field distribution distortion and PD magnitude were used in PD applied voltage (PDAV) and PD inception voltage (PDIV) to evaluate the severity of PD. The ratio (SO2F2)/(SOF2 + SO2) was applied as a criterion to specify the level of PD [51]. Meanwhile, characterising PD severity by using the average magnitude of PD was proposed by [88]. There were three levels assigned for PD severity, namely, mild, medium, and dangerous. The product growth had been correlated with the severity of PD at each level. Meanwhile, the BPNN and SVM algorithms were selected to diagnose it. Hence, it can be concluded that selecting concentration ratios as a feature parameter had shown higher accuracy than the concentration alone to diagnose the PD's severity.
A study [24] used the FCM algorithm to divide the PD into three levels based on the applied voltage, namely slight, medium, and severe, by using seven characteristic quantities of PD as feature parameters before correlating the SF6 byproducts at each severity level. A decision tree algorithm was then applied to obtain the PD status membership function value. A fuzzy comprehensive evaluation theory was finally used to assess the PD status. Previous studies have looked at the concentration of characteristic gases, which varies depending on the SF6 decomposition process. Generally, studies on the severity evaluation criteria are required, and the severity of PD under different configurations of insulation defects needs additional investigation. Therefore, Table V summarises the techniques used to evaluate PD severity in GIS.

VII. DISCUSSION
This study has presented significant drawbacks in the analysis of PD activity in GIE, which are related to PD detection, features parameters selection, and PD data analysis. The following section presents the various factors that could affect the SF6 decomposition process, and hence; the PD diagnosis in general.

A. Comprehensive DCA Based PD Diagnosis Considering the Influential Factors on SF6 Decomposition
Evaluating the internal status of SF6-insulated equipment has been investigated previously under different conditions. Most of the results obtained were limited in discussing a specific facet that might affect SF6 decomposition such as flaws and impurities. However, such studies were helpful to understand and to provide the big picture of PD diagnosis in GIE by considering various factors. Table VI summarises studies devoted to diagnosing SF6 insulated equipment using DCA and illustrates the different factors that influence gas decomposition with key findings obtained at various PD conditions and measurement equipment. The primary decomposition components used for analysing internal deterioration under different combinations of defects are also presented.
The table also reveals that chemical detection techniques were widely adopted to diagnose the internal status of GIE under various PD conditions; mainly, GC and FTIR. Furthermore, most of the studies have considered the SF6 decomposition under PD. The effects of gas pressure, applied voltage, ageing, and impurities such as H2O and O2 have also been investigated. Thus, it can be concluded that considering these factors is significant for a more comprehensive evaluation of GIS's insulation status.

B. Overview of PD Diagnosis and Future Perspectives
This review highlighted many studies on the monitoring and diagnosing of GIS insulation status. These studies included important observations on the recent advances in PD diagnosis using DCA method and AI algorithms. The applied data is diverse according to various factors and approaches used to collect it. Furthermore, most of the previous studies are based on data collected in the laboratory, which is limited by the gas chamber design used to perform the decomposition tests. Accordingly, these works need to be applied for a real-life GIS to be confirmed. More efforts are needed to consider a comprehensive analysis of PD in GIS. Therefore, the diagram shown in Figure 9 illustrates the critical factors affecting the SF6 gas decomposition and the process of monitoring and diagnosing PD in GIS.
The diagram shows that the decomposition process is influenced by physical, chemical, and operational factors, which have been investigated previously (Table VI). Moreover, due to SF6 decomposition process complexity, the penetration of machine learning has been considered a valuable strategy to diagnose the GIS, on top of the chemical analysis of decomposition by-products. Nevertheless, selecting an appropriate model and feature parameters is critical in the diagnostic process. Machine learning techniques and decomposition by-products are promising tools that can be applied to diagnose the internal condition of GIS, with the ability of machine learning techniques to identify boundaries of various conditions numerically. However, selected model should be generalised to avoid under and overfitting problems, and also to deal with data in terms of its dimensionality and scalability. Besides, feature parameters selection is of great importance for the performance and the accuracy of the diagnostic system. This review was intended to provide up-to-date trends in monitoring and diagnosing GIE that uses SF6 as an insulation medium. This study also has offered advanced techniques for detecting and evaluating the severity of PD for researchers. Moreover, it is helpful for them to conduct further research on the topic.

VIII. SUMMARY AND FUTURE WORKS
Many studies on SF6 gas that describe the leading degradation causes, diagnostic methods, and research findings had been previously reported. Nevertheless, a further review on the SF6 decomposition process factors, such as gas pressure, discharge energy, ageing time, insulation defects, and the presence of impurities, could be helpful. This review has revealed some trends in the diagnosis of SF6 insulated equipment, with a special focus on the PD detection using SF6 decomposition byproduct gas analysis. The study on the SF6 decomposition byproducts has attracted many researchers' attention due to its significance and promising future for PD detection. The main findings of this review are summarised as follows: • The positive correlation between SF6 decomposition byproducts and PD activities in a GIS promises a practical method to detect the occurrence of PD and its causes. Key by-product gases and their ratios may be used to optimise diagnosis accuracy. Furthermore, the PD severity level can also be assessed using the same technique. • In developing the PD diagnosis technique based on SF6 decomposition in GIS, it is paramount to consider factors such as the presence of impurities, defect configuration, gas pressure, applied voltage, and ageing time. • Several issues were faced by many PD detection techniques. Among these are the interference faced especially for on-site applications and the limitation in byproduct gas detection range. Due to limitations often faced by any single diagnosis technique, utilisation of combined techniques for PD detection in a GIS may improve the performance of the PD diagnosis and overall GIS monitoring system. • SF6 by-products analyses using machine learning algorithms show a promising detection approach. Nevertheless, feature's parameters and pattern recognition algorithm selection are a challenge that needs to be addressed further, especially in the presence of multiple PD sources.    The ratio (SOF2 + SO2)/SO2F2 GC-MS Corona discharge Point-plane The ratio of (SOF2 +SO2)/SO2F2 decreases with voltage increasing, and increases with SF6 pressure increasing. [89] Effective energy characteristic ratio of (SO2F2)/(SOF2) GC PD Needle-plate electrode Effective characteristics energy ratio proposed based on the ratio (SO2F2)/(SOF2), the higher the ratio is, the larger the PD energy is. [29] PDAV and PDIV GC-MS PD Point-plane electrode Generation rate of SO2F2, SOF2 + SO2, increases with the rise of PDAV, while decreases with the increasing of PDIV [51] Three levels of PD based on PD quantities and decomposition products

GC-MS
Negative DC PD Needle-plate electrode PD status is evaluated based on its association with the generation and concentration of CF4, CO2, SO2F2, SOF2, and SO2) [24] Three levels of PD based on products concentrations GC-MS Negative DC PD Free metal particle Using DCA method shows good performance to diagnose PD severity [88] Effective  [105]