Decomposition Characteristics of SF 6 and Partial Discharge Recognition under Negative DC Conditions

Four typical types of artificial defects are designed in conducting the decomposition experiments of SF6 gas to obtain and understand the decomposition characteristics of SF6 gasinsulated medium under different types of negative DC partial discharge (DC-PD), and use the obtained decomposition characteristics of SF6 in diagnosing the type and severity of insulation fault in DC SF6 gas-insulated equipment. Experimental results show that the negative DC partial discharges caused by the four defects decompose the SF6 gas and generate five stable decomposed components, namely, CF4, CO2, SO2F2, SOF2, and SO2. The concentration, effective formation rate, and concentration ratio of SF6 decomposed components can be associated with the PD types. Furthermore, back propagation neural network algorithm is used to recognize the PD types. The recognition results show that compared with the concentrations of SF6 decomposed components, their concentration ratios are more suitable as the characteristic quantities for PD recognition, and using those concentration ratios in recognizing the PD types can obtain a good effect.


Introduction
Given the connection of new large-scale energy and the rapid development of HVDC transmission technology and flexible HVDC technology, DC gas-insulated equipment (DC-GIE) has attracted significant attention because of its technological advantages in improving system reliability and reducing space occupation [1][2][3][4][5][6].Pure SF 6 gas has stable chemical properties and is not easily decomposed.However, under the effect of the partial discharge (PD), spark discharge, arc discharge, overheating, and other factors, SF 6 gas is decomposed into various low-fluoride sulfides (such as SF 5 , SF 4 , SF 3 , SF 2 , and SF).These low-fluoride sulfides then react with the trace air and moisture that inevitably exist in the DC-GIE to produce stable decomposition products, such as sulfuryl fluoride (SO 2 F 2 ), thionyl fluoride (SOF 2 ), thionyl tetrafluoride (SOF 4 ), sulfur dioxide (SO 2 ), carbon tetrafluoride (CF 4 ), carbon dioxide (CO 2 ), hydrogen fluoride (HF), and hydrogen sulfide (H 2 S) [7][8][9][10][11][12].Some of these products, particularly SO 2 , HF, and H 2 S, can corrode metal parts and solid insulation in equipment, thereby accelerating insulation aging and reducing the overall insulation performance of the equipment.Eventually, the safe and reliable operation of the equipment and the entire power grid is compromised.SF 6 gas component analysis (GCA) is a non-electrical detection method, which can effectively avoid the complex electromagnetic interference of substations and has become a popular research topic in recent years [13][14][15][16].In [17][18][19], the SF 6 decomposed products under arc, spark, and corona discharges were studied.Tang et al. compared and analyzed the decomposition characteristics of SF 6 under two types of common PDs [20], and recognized four typical PD types [21,22].A large number of studies have shown that the type, concentration, and formation law of SF 6 decomposed components are closely related to the type and severity of insulation faults in GIE.However, the recent studies have mainly focused on monitoring the insulation status of AC-GIE using the decomposition characteristics of SF 6 under PD [23][24][25][26][27][28].The research focus has not extended to the DC field yet.
Therefore, four artificial models are designed in this study to simulate the common insulation defects in DC-GIE.These four defect models are placed in the built decomposition device of SF 6 under negative DC conditions, respectively, and PD initial voltage is applied 1.2 times on the defect.As a result, SF 6 decomposes differently under the four defects, and the decomposed components are quantitatively detected by the gas chromatography/mass spectrometry (GC/MS) detection method.The decomposition characteristics of SF 6 under four types of negative DC partial discharge (DC-PD) are obtained.The relationship between the decomposition characteristics of SF 6 and the PD types was further studied.Moreover, back propagation (BP) neural network algorithm [29][30][31][32][33][34][35][36][37][38] is used to recognize the PD types.Then, the best characteristic quantity for PD recognition is extracted.This study lays a solid foundation of using GCA method to diagnose the insulation faults in DC-GIE and assess its insulation status.

Experimental Wiring
The experimental wiring of SF 6 decomposition under negative DC-PD is illustrated in Figure 1.A voltage regulator (T 1 : 0-380 V) and a testing transformer (T 2 : 50 kVA/100 kV) provide the AC high voltage (AC-HV).The AC-HV is converted into DC testing voltage by using a half-wave rectifier circuit, which comprises a HV silicon stack (D s : 100 kV/5 A) and a filter capacitor (C f : 0.2 µF).Two protective resistors (R 1 : 20 kΩ, R 2 : 20 kΩ) are used to protect the system.A capacitive voltage divider (C v ) is used to measure the value of AC output voltage of the transformer, and a resistive voltage divider (R v ) is employed to measure the value of DC testing voltage applied across the defect.A coupling capacitor (C k : 500 pF/100 kV) is used to extract the pulse voltage.A non-inductive detection impedance (Z m : 50 Ω) is used to send the pulse current signal to the digital storage oscilloscope (DSO).The DSO (WavePro 7100XL, New York, USA, analogue band: 1 GHz, sampling rate: 20 GHz, memory depth: 48 MB) is used to monitor the PD magnitude.The structure of the gas chamber is shown in Figure 2, and its volume is 60 L. A GC/MS (Shimadzu QP-2010Ultra, Kyoto, Japan, precision: 0.01 ppm, accuracy: ±10%) is used to measure the sample gas components quantitatively.The GC/MS uses He (purity: 99.9995%) as the carrier gas and deploys the special capillary column CP-Sil5CB (Shimadzu, Kyoto, Japan) to separate different components.
topic in recent years [13][14][15][16].In [17][18][19], the SF6 decomposed products under arc, spark, and corona discharges were studied.Tang et al. compared and analyzed the decomposition characteristics of SF6 under two types of common PDs [20], and recognized four typical PD types [21,22].A large number of studies have shown that the type, concentration, and formation law of SF6 decomposed components are closely related to the type and severity of insulation faults in GIE.However, the recent studies have mainly focused on monitoring the insulation status of AC-GIE using the decomposition characteristics of SF6 under PD [23][24][25][26][27][28].The research focus has not extended to the DC field yet.
Therefore, four artificial models are designed in this study to simulate the common insulation defects in DC-GIE.These four defect models are placed in the built decomposition device of SF6 under negative DC conditions, respectively, and PD initial voltage is applied 1.2 times on the defect.As a result, SF6 decomposes differently under the four defects, and the decomposed components are quantitatively detected by the gas chromatography/mass spectrometry (GC/MS) detection method.The decomposition characteristics of SF6 under four types of negative DC partial discharge (DC-PD) are obtained.The relationship between the decomposition characteristics of SF6 and the PD types was further studied.Moreover, back propagation (BP) neural network algorithm [29][30][31][32][33][34][35][36][37][38] is used to recognize the PD types.Then, the best characteristic quantity for PD recognition is extracted.This study lays a solid foundation of using GCA method to diagnose the insulation faults in DC-GIE and assess its insulation status.

Experimental Wiring
The experimental wiring of SF6 decomposition under negative DC-PD is illustrated in Figure 1.A voltage regulator (T1: 0-380 V) and a testing transformer (T2: 50 kVA/100 kV) provide the AC high voltage (AC-HV).The AC-HV is converted into DC testing voltage by using a half-wave rectifier circuit, which comprises a HV silicon stack (Ds: 100 kV/5 A) and a filter capacitor (Cf: 0.2 μF).Two protective resistors (R1: 20 kΩ, R2: 20 kΩ) are used to protect the system.A capacitive voltage divider (Cv) is used to measure the value of AC output voltage of the transformer, and a resistive voltage divider (Rv) is employed to measure the value of DC testing voltage applied across the defect.A coupling capacitor (Ck: 500 pF/100 kV) is used to extract the pulse voltage.A non-inductive detection impedance (Zm: 50 Ω) is used to send the pulse current signal to the digital storage oscilloscope (DSO).The DSO (WavePro 7100XL, New York, USA, analogue band: 1 GHz, sampling rate: 20 GHz, memory depth: 48 MB) is used to monitor the PD magnitude.The structure of the gas chamber is shown in Figure 2, and its volume is 60 L. A GC/MS (Shimadzu QP-2010Ultra, Kyoto, Japan, precision: 0.01 ppm, accuracy: ±10%) is used to measure the sample gas components quantitatively.The GC/MS uses He (purity: 99.9995%) as the carrier gas and deploys the special capillary column CP-Sil5CB (Shimadzu, Kyoto, Japan) to separate different components.

Insulation Defect Model
As shown in Figure 3, typical defects in practical DC-GIE include: metallic protrusion, which is usually manifested as the abnormal bulge on a HV conductor [39]; free conductive particle, which is generally shown as the metal powder that can move freely in a cavity [40]; insulator pollution, which is formed by various pollution on the surface of an insulator [41]; and insulator gap, which is formed by peeling a gap between a HV conductor and disc insulator [42].The protrusion defect is the most harmful and the particle defect is the most typical in DC-GIE.According to the characteristics of these defects, the four defect models (protrusion defect, particle defect, pollution defect, and gap defect) are designed in this study for experimental research, as shown in Figure 4.All electrodes shown in Figure 4 are made of stainless steel, and all plate electrodes have the same size: the thickness and diameter are 10 mm and 120 mm, respectively.The experiment in this study uses a needle-plate model to simulate the protrusion defect (Figure 4a).The distance between the needle and the plate is 10 mm.The curvature radius of the needle tip is 0.3 mm.The model of the particle defect consists of a ball electrode (HV electrode), a bowl electrode (ground electrode), and 20 aluminum balls (Figure 4b).

Insulation Defect Model
As shown in Figure 3, typical defects in practical DC-GIE include: metallic protrusion, which is usually manifested as the abnormal bulge on a HV conductor [39]; free conductive particle, which is generally shown as the metal powder that can move freely in a cavity [40]; insulator pollution, which is formed by various pollution on the surface of an insulator [41]; and insulator gap, which is formed by peeling a gap between a HV conductor and disc insulator [42].The protrusion defect is the most harmful and the particle defect is the most typical in DC-GIE.According to the characteristics of these defects, the four defect models (protrusion defect, particle defect, pollution defect, and gap defect) are designed in this study for experimental research, as shown in Figure 4.All electrodes shown in Figure 4 are made of stainless steel, and all plate electrodes have the same size: the thickness and diameter are 10 mm and 120 mm, respectively.The experiment in this study uses a needle-plate model to simulate the protrusion defect (Figure 4a).The distance between the needle and the plate is 10 mm.The curvature radius of the needle tip is 0.3 mm.The model of the particle defect consists of a ball electrode (HV electrode), a bowl electrode (ground electrode), and 20 aluminum balls (Figure 4b).

Insulation Defect Model
As shown in Figure 3, typical defects in practical DC-GIE include: metallic protrusion, which is usually manifested as the abnormal bulge on a HV conductor [39]; free conductive particle, which is generally shown as the metal powder that can move freely in a cavity [40]; insulator pollution, which is formed by various pollution on the surface of an insulator [41]; and insulator gap, which is formed by peeling a gap between a HV conductor and disc insulator [42].The protrusion defect is the most harmful and the particle defect is the most typical in DC-GIE.According to the characteristics of these defects, the four defect models (protrusion defect, particle defect, pollution defect, and gap defect) are designed in this study for experimental research, as shown in Figure 4.All electrodes shown in Figure 4 are made of stainless steel, and all plate electrodes have the same size: the thickness and diameter are 10 mm and 120 mm, respectively.The experiment in this study uses a needle-plate model to simulate the protrusion defect (Figure 4a).The distance between the needle and the plate is 10 mm.The curvature radius of the needle tip is 0.3 mm.The model of the particle defect consists of a ball electrode (HV electrode), a bowl electrode (ground electrode), and 20 aluminum balls (Figure 4b).The diameter of the ball electrode is 50 mm.A hollow sphere is cut to obtain the bowl electrode.The diameters of the hollow sphere, circular incision, and aluminum ball are 100 mm, 90 mm, and 3 mm, respectively.The ball and bowl electrodes form a concentric sphere structure.Some copper cuttings are adhered on the surface of an epoxy cylinder to simulate the pollution defect (Figure 4c).The epoxy is inserted between two plate electrodes.The size of the copper cuttings is 4 mm × 20 mm.The epoxy cylinder is 25 mm thick and 80 mm in diameter.An epoxy cylinder is inserted between the two plate electrodes to create the gap defect model (Figure 4d).A 2-mm gap is maintained between the HV electrode and the epoxy.The epoxy cylinder is 50 mm thick and 80 mm in diameter.

Experimental Method
The experimental system is connected as shown in Figure 1.The defect model is placed in the gas chamber, which is cleaned three times with SF6 gas.Then, the chamber is filled with 0.2 MPa SF6.The concentration of H2O and O2 in the chamber satisfy the industrial standard of DL/T 596-1996 [43].The experimental voltage is raised gradually until the oscilloscope can detect the PD signal.This voltage is the PD initial voltage (U0).The decomposition experiment of SF6 is conducted for 96 h under the voltage of 1.2 U0.The SF6 decomposed components are collected every 12 h.GC/MS is used to measure the component concentration.All the experiments are conducted under the same conditions to ensure comparability.The laboratory temperature and relative humidity are maintained at 20 °C and 50%, respectively.The purity of SF6 gas is 99.9995%.
The experimental results show that the PD initial voltage of the system without the defect is 82 kV.The PD initial voltages of the system after placing the four defects, and the corresponding experimental voltages are shown in Table 1.No breakdown occurred at each experimental voltage.The diameter of the ball electrode is 50 mm.A hollow sphere is cut to obtain the bowl electrode.The diameters of the hollow sphere, circular incision, and aluminum ball are 100 mm, 90 mm, and 3 mm, respectively.The ball and bowl electrodes form a concentric sphere structure.Some copper cuttings are adhered on the surface of an epoxy cylinder to simulate the pollution defect (Figure 4c).The epoxy is inserted between two plate electrodes.The size of the copper cuttings is 4 mm × 20 mm.The epoxy cylinder is 25 mm thick and 80 mm in diameter.An epoxy cylinder is inserted between the two plate electrodes to create the gap defect model (Figure 4d).A 2-mm gap is maintained between the HV electrode and the epoxy.The epoxy cylinder is 50 mm thick and 80 mm in diameter.

Experimental Method
The experimental system is connected as shown in Figure 1.The defect model is placed in the gas chamber, which is cleaned three times with SF 6 gas.Then, the chamber is filled with 0.2 MPa SF 6 .The concentration of H 2 O and O 2 in the chamber satisfy the industrial standard of DL/T 596-1996 [43].The experimental voltage is raised gradually until the oscilloscope can detect the PD signal.This voltage is the PD initial voltage (U 0 ).The decomposition experiment of SF 6 is conducted for 96 h under the voltage of 1.2 U 0 .The SF 6 decomposed components are collected every 12 h.GC/MS is used to measure the component concentration.All the experiments are conducted under the same conditions to ensure comparability.The laboratory temperature and relative humidity are maintained at 20 • C and 50%, respectively.The purity of SF 6 gas is 99.9995%.
The experimental results show that the PD initial voltage of the system without the defect is 82 kV.The PD initial voltages of the system after placing the four defects, and the corresponding experimental voltages are shown in Table 1.No breakdown occurred at each experimental voltage.

PD Characteristics
PD has a cumulative effect on SF 6 decomposition.That is, PD repetition rate (N, unit: pulse/s), the average discharge magnitude of a single pulse (Q avg , unit: pC/pulse), and the average discharge magnitude in a second (Q sec , unit: pC/s) affect the SF 6 decomposition.A certain relationship exists among these three parameters (N, Q avg , and Q sec ), as shown in Equation (1).The measurements of N, Q avg , and Q sec under the four defects are shown in Table 2.This experiment uses a 50 Ω non-inductive resistor pulse detection unit to monitor the PD waveforms.A PD calibration circuit is used to calibrate the PD magnitude, as recommended by IEC 60270:2000 [44]: Table 2 shows that the PD characteristics under the four defects are significantly different.Under the protrusion defect, Q avg is the smallest, N and Q sec are the largest, and discharge is the most stable.In this study, the aluminum balls are selected as the metal particles that can move freely.The aluminum balls easily jump under the influence of an electric field.After the aluminum balls fall, they slide to the bottom of the bowl electrode.These balls have a concentrated distribution while moving.Therefore, the PD under the particle defect is stable.Moreover, the PD has the second largest N, Q avg , and Q sec under the particle defect compared with those under the other three defects.Under the pollution defect, Q avg and Q sec are small, and N is slightly smaller than that under the particle defect.Under the same experimental conditions, the gap defect has the most difficulty in producing PD, under this defect, Q avg is the largest, N and Q sec are the smallest, and discharge has a large dispersion.In summary, the N, Q avg , and Q sec under the four defects are significantly different, which is the basic reason for the difference in the SF 6 decomposition.This result provides a possibility to study and establish the corresponding relationship between the decomposition characteristics of SF 6 and the defect types.

Concentrations of SF 6 Decomposed Components
Experimental results show that the negative DC partial discharges caused by the four defects decompose SF 6 gas and generate five stable decomposed components, namely, CF 4 , CO 2 , SO 2 F 2 , SOF 2 , and SO 2 .The relationship between these five decomposed components and the defect types is investigated in this study.

Concentrations of CF 4 and CO 2
As shown in Figure 5a, the CF 4 concentration presents a "linearly saturated" growth trend with time under the particle defect.Under the protrusion and pollution defects, the CF 4 concentrations present an approximately linear growth trend with time.Under the gap defect, the CF 4 concentration features the popular logistic population growth trend called the "S" growth trend with time [45,46].At 96 h, the CF 4 concentration is 2.37 ppm under the particle defect.The CF 4 concentrations are less than 0.32 ppm under the other three defects.Thus, the generation amounts of CF 4 are small, and the change curves of the CF 4 concentrations under these three defects are close to one another and have two intersections in 24 h.Errors existed when GC/MS is used to detect the gas concentration.Therefore, this study does not recommend selecting CF 4 concentration as a characteristic quantity to recognize the PD types.However, CF 4 concentration can be used as an auxiliary criterion for identifying the particle defect.
Energies 2017, 10, 556 6 of 16 features the popular logistic population growth trend called the "S" growth trend with time [45,46].At 96 h, the CF4 concentration is 2.37 ppm under the particle defect.The CF4 concentrations are less than 0.32 ppm under the other three defects.Thus, the generation amounts of CF4 are small, and the change curves of the CF4 concentrations under these three defects are close to one another and have two intersections in 24 h.Errors existed when GC/MS is used to detect the gas concentration.Therefore, this study does not recommend selecting CF4 concentration as a characteristic quantity to recognize the PD types.However, CF4 concentration can be used as an auxiliary criterion for identifying the particle defect.As shown in Figure 5b, the CO2 concentration presents an approximately linear growth trend with time under the gap defect.Under the other three defects, the CO2 concentrations present a "linearly saturated" growth trend with time.At 96 h, the concentration relationship of CO2 under the four defects is as follows: protrusion defect (20.3 ppm) > pollution defect (9.77 ppm) > particle defect (5.67 ppm) > gap defect (4.69 ppm).CO2 is generated by the reaction of C atoms and O2.Under the effect of the electric field, the charged particles hit the surface of the electrodes and release C atoms.During the experiment, the applied voltage remains constant, and C atoms are produced continually.Therefore, the growth trend of CO2 is mainly determined by the O2 concentration in the chamber.In the early stage of the experiment, O2 is abundant, thus, the CO2 concentrations increase linearly with time under the four defects.Under the gap defect, the N and Qsec are small, and the generation amounts of SF6 decomposed components are small.Hence, the consumed O2 concentration is also small.In the late stage of the experiment, O2 is also abundant, thus, the CO2 concentration increases linearly with time under the gap defect.Compared with those under the gap defect, the generation amounts of SF6 decomposed components under the other three defects are significantly larger, thus, the consumed O2 concentrations are also remarkably larger.In the late stage of the experiment, the remaining small amount of O2 inhibits the CO2 formation to some extent, thus, the CO2 concentrations present a saturated growth trend with time under the other three defects.According to the growth As shown in Figure 5b, the CO 2 concentration presents an approximately linear growth trend with time under the gap defect.Under the other three defects, the CO 2 concentrations present a "linearly saturated" growth trend with time.At 96 h, the concentration relationship of CO 2 under the four defects is as follows: protrusion defect (20.3 ppm) > pollution defect (9.77 ppm) > particle defect (5.67 ppm) > gap defect (4.69 ppm).CO 2 is generated by the reaction of C atoms and O 2 .Under the effect of the electric field, the charged particles hit the surface of the electrodes and release C atoms.During the experiment, the applied voltage remains constant, and C atoms are produced continually.Therefore, the growth trend of CO 2 is mainly determined by the O 2 concentration in the chamber.In the early stage of the experiment, O 2 is abundant, thus, the CO 2 concentrations increase linearly with time under the four defects.Under the gap defect, the N and Q sec are small, and the generation amounts of SF 6 decomposed components are small.Hence, the consumed O 2 concentration is also small.In the late stage of the experiment, O 2 is also abundant, thus, the CO 2 concentration increases linearly with time under the gap defect.Compared with those under the gap defect, the generation amounts of SF 6 decomposed components under the other three defects are significantly larger, thus, the consumed O 2 concentrations are also remarkably larger.In the late stage of the experiment, the remaining small amount of O 2 inhibits the CO 2 formation to some extent, thus, the CO 2 concentrations present a saturated growth trend with time under the other three defects.According to the growth trend and the value of CO 2 concentration, the four defects can be distinguished from one another.Therefore, CO 2 concentration can be selected as a characteristic quantity to recognize the PD types.
3.2.2.Concentrations of SO 2 F 2 , SOF 2 , and SO 2 Under the four defects, the concentrations of SO 2 F 2 , SOF 2 , and SO 2 approximately present a "linearly saturated" growth trend with time (Figure 6).trend and the value of CO2 concentration, the four defects can be distinguished from one another.Therefore, CO2 concentration can be selected as a characteristic quantity to recognize the PD types.

Concentrations of SO2F2, SOF2, and SO2
Under the four defects, the concentrations of SO2F2, SOF2, and SO2 approximately present a "linearly saturated" growth trend with time (Figure 6).The concentration relationships of SO2F2, SOF2, and SO2 under the four defects are the same, namely, protrusion defect > particle defect > pollution defect > gap defect, which are consistent with the value relationships of N and Qsec under the four defects.Thus, the N and Qsec of PD are the main factors that influence the formations of SO2F2, SOF2, and SO2.The differences among the curves in Figure 6a,b are evident.According to the growth trend and the value of SO2 concentration, the four The concentration relationships of SO 2 F 2 , SOF 2 , and SO 2 under the four defects are the same, namely, protrusion defect > particle defect > pollution defect > gap defect, which are consistent with the value relationships of N and Q sec under the four defects.Thus, the N and Q sec of PD are the main factors that influence the formations of SO 2 F 2 , SOF 2 , and SO 2 .The differences among the curves in Figure 6a,b are evident.According to the growth trend and the value of SO 2 concentration, the four defects can be distinguished from one another.Therefore, the concentrations of SO 2 F 2 , SOF 2 , and SO 2 can be selected as characteristic quantities to recognize the PD types.
The formation processes of SO 2 F 2 and SOF 2 are shown in Figure 7.Under the effect of the electric field, the charged particles hit the SF 6 molecule and generate low-fluoride sulfides (such as SF 5  defects can be distinguished from one another.Therefore, the concentrations of SO2F2, SOF2, and SO2 can be selected as characteristic quantities to recognize the PD types.The formation processes of SO2F2 and SOF2 are shown in Figure 7.Under the effect of the electric field, the charged particles hit the SF6 molecule and generate low-fluoride sulfides (such as SF5, SF4, SF3, and SF2), and hit the H2O and O2 molecules and generate O atom and OH ion.During the experiment, the applied voltage remains constant, and low-fluoride sulfides are produced continually.Therefore, the growth trends of SO2F2 and SOF2 are mainly determined by the concentrations of H2O and O2 in the chamber.In the early stage of the experiment, H2O and O2 are abundant, thus, the concentrations of SO2F2 and SOF2 increase linearly with time.H2O and O2 in the chamber are gradually consumed with time.Moreover, the formation rates of SO2F2 and SOF2 are significantly larger than other components.In the late stage of the experiment, the further increase in the concentrations of SO2F2 and SOF2 at large rates is difficult to ensure with the remaining H2O and O2 in the chamber, thus the formations of SO2F2 and SOF2 are prevented to some extent.Therefore, the concentrations of SO2F2 and SOF2 present a saturated growth trend with time in the late stage of the experiment.SO2 is produced by the hydrolysis reaction of SOF2.Thus, the growth trend of SO2 is determined by the concentrations of SOF2 and H2O.The SOF2 concentrations present a "linearly saturated" growth trend with time under the four defects.Similarly, the SO2 concentrations also present a "linear-saturated" growth trend with time under the four defects because of the growth trend of SOF2 concentrations and the consumption condition of H2O in the chamber.

Formation Rates of SF6 Decomposed Components
The severity of the insulation defect is difficult to assess accurately based only on the concentrations of SF6 decomposed components.The formation rates of SF6 decomposed components should also be considered because they can directly reflect the fault consumed energy, fault type, fault severity, and fault development process.Therefore, this study adopts the effective formation rate (RRMS) [47] to study the relationship between SF6 decomposed components and PD types.RRMS is expressed as follows: where Raij is the absolute formation rate of component i in day j.Ra is the absolute formation rate (ppm/day) and is calculated using the following equation: where Ci1 is the concentration of component i from the first measurement time, Ci2 is the concentration of component i from the second measurement time, △t is the time interval between two measurement times, and △t = 1 day in this study.SO 2 is produced by the hydrolysis reaction of SOF 2 .Thus, the growth trend of SO 2 is determined by the concentrations of SOF 2 and H 2 O.The SOF 2 concentrations present a "linearly saturated" growth trend with time under the four defects.Similarly, the SO 2 concentrations also present a "linear-saturated" growth trend with time under the four defects because of the growth trend of SOF 2 concentrations and the consumption condition of H 2 O in the chamber.

Formation Rates of SF 6 Decomposed Components
The severity of the insulation defect is difficult to assess accurately based only on the concentrations of SF 6 decomposed components.The formation rates of SF 6 decomposed components should also be considered because they can directly reflect the fault consumed energy, fault type, fault severity, and fault development process.Therefore, this study adopts the effective formation rate (R RMS ) [47] to study the relationship between SF 6 decomposed components and PD types.R RMS is expressed as follows: where R aij is the absolute formation rate of component i in day j.R a is the absolute formation rate (ppm/day) and is calculated using the following equation: where C i1 is the concentration of component i from the first measurement time, C i2 is the concentration of component i from the second measurement time, t is the time interval between two measurement times, and t = 1 day in this study.The effective formation rates of SF 6 decomposed components under the four defects are shown in Table 3.Under the protrusion defect, the R RMS of CF 4 is the smallest, and the effective formation rates of CO 2 , SO 2 F 2 , SOF 2 , and SO 2 are remarkably larger than those under the other three defects.The R RMS of SOF 2 under the particle defect is three times the R RMS of SOF 2 under the pollution defect, which is eight times the R RMS of SOF 2 under the gap defect.The R RMS of CO 2 under the pollution defect is nearly two times the effective formation rates of CO 2 under the particle and gap defects.Under the gap defect, the sum of the effective formation rates of SF 6 decomposed components is the smallest.The effective formation rates of SF 6 decomposed components under the four defects are different from one another, thus, they can be used to recognize the PD types.Under the particle defect, the aluminum balls (Weihua Aluminum Industry Co., Ltd., Wuhan, China, mass fraction of C element: 0.88%) can provide C atoms for CF 4 formation, and the PD has the second largest N, Q avg , and Q sec , thus, the R RMS of CF 4 is the largest.Under the pollution defect, the epoxy block can provide abundant C atoms for CO 2 formation, thus, the R RMS of CO 2 is larger than that under the particle defect.Under the protrusion, pollution, and gap defects, the R RMS relationship of CO 2 is consistent with the value relationships of N and Q sec .Therefore, the R RMS of CO 2 is determined by the following two aspects: (1) the N and Q sec of PD; (2) whether an organic insulating material is present near the PD area.Under the four defects, the R RMS relationships of SO 2 F 2 , SOF 2 , and SO 2 are consistent with the value relationships of N and Q sec .Thus, the N and Q sec of PD are the main factors that influence the effective formation rates of SO 2 F 2 , SOF 2 , and SO 2 .

Concentration Ratios of SF 6 Decomposed Components
The gas chamber volume and trace H 2 O and O 2 affect the concentrations and formation rates of SF 6 decomposed components.Based on the three-ratio method in oil chromatographic analysis, C(SOF 2 )/C(SO 2 F 2 ), C(CF 4 )/C(CO 2 ), and C(SO 2 F 2 + SOF 2 )/C(CF 4 + CO 2 ) were selected as characteristic quantities to recognize the PD types [21,22].These three concentration ratios have definite physical meaning.C(SOF 2 )/C(SO 2 F 2 ) is used to describe PD severity; C(CF 4 )/C(CO 2 ) is used to describe the structure of insulation defect; C(SO 2 F 2 + SOF 2 )/C(CF 4 + CO 2 ) is used to describe the deterioration degree of carbonaceous material (the metal part and organic insulation).In [21,22], the formation characteristics of SO 2 were not studied because of the limitations of the experimental conditions.SOF 2 is prone to hydrolysis in producing SO 2 .Currently, the common electrochemical sensors can respond to SOF 2 and SO 2 in the field overhaul.IEC 60480:2004 [48] stipulates that the sum of the concentrations of SOF 2 and SO 2 in the SF 6 gas that can be recycled cannot exceed 12 ppm.Thus, SO 2 is an important component and can be selected as a characteristic gas for field detection.Based on the experimental data in this study, C(SOF 2 + SO 2 )/C(SO 2 F 2 ), Ln(C(CO 2 )/C(CF 4 )), and C(SO 2 F 2 + SOF 2 + SO 2 )/C(CF 4 + CO 2 ) are selected as characteristic quantities to study the relationship between the decomposition characteristics of SF 6 and PD types.
The value relationship of C(SOF 2 + SO 2 )/C(SO 2 F 2 ) under the four defects is: particle defect > protrusion defect > pollution defect > gap defect (Figure 8a).A smaller value of C(SOF 2 + SO 2 )/C(SO 2 F 2 ) indicates a larger PD energy and a more serious fault.Therefore, the PD energy under the particle defect is the smallest, and the PD energies under the other three defects are close to one another.The curves in Figure 8b can be easily distinguished from one another, the curves have no intersection, indicating significant differences among the structures of the four defects.Thus, Ln(C(CO 2 )/C(CF 4 )) can be used to recognize the four defects in this study.A smaller value of C(SO2F2 + SOF2 + SO2)/C(CF4 + CO2) indicates a more serious deterioration of the carbonaceous materials.Therefore, the deteriorations of the carbonaceous materials under the pollution and gap defects are the most serious, and that under the protrusion defect is the lightest.
If the concentration ratios of the SF6 decomposed components are to be used in recognizing the PD types in the field fault diagnosis, then the selected ratios cannot present significant fluctuations with time.Even if insulation defects exist in GIE, the concentrations of SF6 decomposed components have been basically stable after a long period of operation, thus, the concentration ratios are unlikely to appear as large fluctuations with time.As shown in Figure 8, the values of C(SOF2 + SO2)/C(SO2F2), A smaller value of C(SO 2 F 2 + SOF 2 + SO 2 )/C(CF 4 + CO 2 ) indicates a more serious deterioration of the carbonaceous materials.Therefore, the deteriorations of the carbonaceous materials under the pollution and gap defects are the most serious, and that under the protrusion defect is the lightest.
If the concentration ratios of the SF 6 decomposed components are to be used in recognizing the PD types in the field fault diagnosis, then the selected ratios cannot present significant fluctuations with time.Even if insulation defects exist in GIE, the concentrations of SF 6 decomposed components have been basically stable after a long period of operation, thus, the concentration ratios are unlikely to appear as large fluctuations with time.As shown in Figure 8

PD Recognition
This study uses the back propagation (BP) neural network algorithm [29][30][31][32][33][34][35][36][37][38] to recognize the PD types.The concentrations and concentration ratios of SF 6 decomposed components are selected as the input matrix of the network.Then the recognition results are analyzed, and the best characteristic quantity for PD recognition is extracted.
BP neural network is a multi-layer feedforward network, which uses the steepest descent method to achieve the minimum of mean square error between the expected output and actual output.The structure of BP neural network is shown in Figure 9, including the input, hidden, and output layers.The learning process of BP neural network is composed of the forward propagation (FP) of signal and the BP of error.In the FP process, the input signals are transferred from the input layer to the hidden layer until the output layer.If the output layer does not obtain an expected output, then the weights and thresholds should be adjusted to reduce the predicted error in the BP process.

PD Recognition
This study uses the back propagation (BP) neural network algorithm [29][30][31][32][33][34][35][36][37][38] to recognize the PD types.The concentrations and concentration ratios of SF6 decomposed components are selected as the input matrix of the network.Then the recognition results are analyzed, and the best characteristic quantity for PD recognition is extracted.
BP neural network is a multi-layer feedforward network, which uses the steepest descent method to achieve the minimum of mean square error between the expected output and actual output.The structure of BP neural network is shown in Figure 9, including the input, hidden, and output layers.The learning process of BP neural network is composed of the forward propagation (FP) of signal and the BP of error.In the FP process, the input signals are transferred from the input layer to the hidden layer until the output layer.If the output layer does not obtain an expected output, then the weights and thresholds should be adjusted to reduce the predicted error in the BP process.  5.
Step 2: Initialize the network.Random numbers in the range of [0, 1] are assigned to the connection weights and the thresholds.Set the maximum number of iterations is 100, the learning rate is 0.1, the permissible error between expected output and actual output is 0.01, and the number of the hidden layers is 10.
Step 3: Input the training samples.The 75% (18 groups) of the above 24 groups of experimental data are randomly extracted as the training samples, and the remaining 25% data (6 groups) are used as the test samples.The training samples are used to train the BP neural network model.Then, the trained model is used to recognize the test samples.
Step 4: Train BP neural network and recognize PD types.To make the recognition results  5.
Step 2: Initialize the network.Random numbers in the range of [0, 1] are assigned to the connection weights and the thresholds.Set the maximum number of iterations is 100, the learning rate is 0.1, the permissible error between expected output and actual output is 0.01, and the number of the hidden layers is 10.
Step 3: Input the training samples.The 75% (18 groups) of the above 24 groups of experimental data are randomly extracted as the training samples, and the remaining 25% data (6 groups) are used as the test samples.The training samples are used to train the BP neural network model.Then, the trained model is used to recognize the test samples.
Step 4: Train BP neural network and recognize PD types.To make the recognition results statistically significant, Step 3 is repeated 100 times.Thus, the number of the test samples is 600.The trained BP neural network models are used to recognize these 600 groups of test samples, the recognition results are shown in Table 6.Therefore, the concentration ratios of SF 6 decomposed components are used to recognize the PD types, the total recognition accuracy rate is: (169 + 132 + 114 + 109)/600 ≈ 87%.Similarly, the concentrations of SF 6 decomposed components are selected as the input matrix of the network.From the previous analysis, CF 4 concentration is not suitable as a characteristic quantity to recognize the PD types.Thus, the input matrix is where k = 1, 2, . . ., 32.The output matrix remains unchanged (Table 5).Steps 2-4 are conducted, and the recognition results are shown in Table 7.Therefore, the concentrations of SF 6 decomposed components are used to recognize PD types, the total recognition accuracy rate is: (167 + 165 + 105 + 104)/800 ≈ 68%.Evidently, the total accuracy rate of PD recognition by using the concentration ratios is nearly 20% higher than that by using the concentrations.Moreover, c(SOF 2 + SO 2 )/c(SO 2 F 2 ), Ln(c(CO 2 )/c(CF 4 )), and c(SOF 2 + SO 2 F 2 + SO 2 )/c(CF 4 + CO 2 ) have definite physical meaning.Therefore, these three concentration ratios are more suitable as the characteristic quantities for PD recognition than the concentrations of SF 6 decomposed components.
To test the recognition performance of BP neural network, the 24 groups of concentration ratios data in this study are selected to train the network model.Then, the trained model is used to recognize another 24 groups of concentration ratio data produced by the same experiment.These 24 groups of data are also basically stable.The deviations of the concentrations of SF 6 decomposed components between these two experiments are less than 10%.The confusion matrix of the recognition result is shown in Table 8, the total recognition accuracy rate is 87.5%, and a good recognition effect is obtained.

Discussion
This research studied the decomposition characteristics of SF 6 under negative DC-PD, and used BP neural network algorithm to recognize four typical insulation faults in DC-GIE.This study could lay a solid foundation of using GCA method to diagnose the insulation faults in DC-GIE and assess its insulation status.However, the concentrations of SF 6 decomposed components are not only related to PD type, they are also affected by PD strength [49], the H 2 O [50], O 2 [51], and absorbent [52] in DC-GIE, and so on.This study did not consider these impact factors.To achieve better performance in using GCA method for PD recognition, the influence of these factors on the concentrations of SF 6 decomposed components must be studied.Moreover, this study used the decomposition characteristics of SF 6 to recognize the four common PD types in DC-GIE, and obtained a good recognition effect in the laboratory.We should focus on the research of the field application in the future, and use the engineering data to verify the validity of the method in this study and make corresponding improvements to this method.

Conclusions
The decomposition characteristics of SF 6 under four types of negative DC partial discharges are obtained in this study.The relationship between the decomposition characteristics of SF 6 and the PD types was further studied.Moreover, BP neural network algorithm is used to recognize the PD types.The following conclusions can be drawn from this study:

•
The negative DC partial discharges caused by the four defects decompose the SF 6 gas and generate five stable decomposed components, namely, CF 4 , CO 2 , SO 2 F 2 , SOF 2 , and SO 2 .A close relationship exists between the decomposition characteristics of SF 6 and the types of insulation defects.The decomposition characteristics of SF 6 can be used to diagnose the type and severity of insulation fault in DC-GIE.

•
BP neural network algorithm is used to recognize the PD types.The recognition results show that the total recognition accuracy rate is 67.63% and 87.33% when the concentrations and concentration ratios of SF 6 decomposed components are selected as the input matrix of the network, respectively.Therefore, the concentration ratios of SF 6 decomposed components are more suitable as the characteristic quantities for PD recognition than the concentrations of those.

Figure 1 .
Figure 1.Experimental wiring of SF6 decomposition under negative DC-PD.Figure 1. Experimental wiring of SF 6 decomposition under negative DC-PD.

Figure 1 .
Figure 1.Experimental wiring of SF6 decomposition under negative DC-PD.Figure 1. Experimental wiring of SF 6 decomposition under negative DC-PD.

Figure 5 .
Figure 5. Change curves of the concentrations of CF4 and CO2 with time: (a) CF4; and (b) CO2.

Figure 5 .
Figure 5. Change curves of the concentrations of CF 4 and CO 2 with time: (a) CF 4 ; and (b) CO 2 .

Figure 9 .
Figure 9. Structure of BP neural network.

Figure 9 .
Figure 9. Structure of BP neural network.

Table 1 .
Experimental voltages under four kinds of defects.

Table 2 .
N, Q avg , and Q sec of PD caused by four kinds of defects.
, SF 4 , SF 3 , and SF 2 ), and hit the H 2 O and O 2 molecules and generate O atom and OH ion.During the experiment, the applied voltage remains constant, and low-fluoride sulfides are produced continually.Therefore, the growth trends of SO 2 F 2 and SOF 2 are mainly determined by the concentrations of H 2 O and O 2 in the chamber.In the early stage of the experiment, H 2 O and O 2 are abundant, thus, the concentrations of SO 2 F 2 and SOF 2 increase linearly with time.H 2 O and O 2 in the chamber are gradually consumed with time.Moreover, the formation rates of SO 2 F 2 and SOF 2 are significantly larger than other components.In the late stage of the experiment, the further increase in the concentrations of SO 2 F 2 and SOF 2 at large rates is difficult to ensure with the remaining H 2 O and O 2 in the chamber, thus the formations of SO 2 F 2 and SOF 2 are prevented to some extent.Therefore, the concentrations of SO 2 F 2 and SOF 2 present a saturated growth trend with time in the late stage of the experiment.

Table 3 .
Effective formation rates of SF 6 decomposed components.

Table 4 .
Input matrix of BP neural network.

Table 5 .
Output matrix of BP neural network.

Table 6 .
Recognition result using the concentration ratios of SF 6 decomposed components as the input matrix.

Table 7 .
Recognition result using the concentrations of SF 6 decomposed components as the input matrix.

Table 8 .
Confusion matrix of the recognition result.
• C(SOF 2 + SO 2 )/C(SO 2 F 2 ), Ln(C(CO 2 )/C(CF 4 )), and C(SO 2 F 2 + SOF 2 + SO 2 )/C(CF 4 + CO 2 ) are used to recognize the PD types.The 24 groups of concentration ratio data in this study are selected to train the BP neural network model.Then, the trained model is used to recognize another 24 groups of concentration ratio data produced by the same experiment.The total recognition accuracy rate is 87.5%, and a good recognition effect is obtained.