High-Frequency Current Transformer Design and Implementation Considerations for Wideband Partial Discharge Applications

High-frequency current transformers are popular noninvasive inductive wideband sensors. Despite the simplicity in design and operational principle, implementation of such sensors for partial discharge (PD) applications requires careful consideration, particularly in the higher frequency range where traveling wave attenuation and distortion are relevant. First, the role of design variables, including core materials, winding design, and shielding practices on sensor sensitivity and frequency characteristics (transfer impedance), is presented. Next, the suitability of the constructed sensors for PD applications is assessed. The designed wideband sensors are suitable for laboratory applications with standard measurement circuits and controlled conditions. The low-level magnitude and frequency spectrum of the discharge pulses hinders signal integrity in relation to the placement of the sensors within the measurement circuit, signal amplification, and pulse repetition rate (pulse resolution). To enable the most stringent detection levels under 1 pC, efforts are needed in distortionless amplifier design and interference mitigation.


I. INTRODUCTION
P ARTIAL discharges (PD) are of great interest to widely ranging research and application themes. In general, one strives to collect representative data of the PD event that enables reliable interpretation of the system, device, or material status. However, the physical PD occurrence cannot be directly measured without altering the event. As illustrated in Fig. 1, PD manifests itself in numerous forms. Charge migration during the discharge event can produce electrical current pulses that can be observed using appropriate instruments; resulting electromagnetic radiation can be detected with Manuscript received August 19, 2020; revised December 7, 2020; accepted December 31, 2020. Date of publication January 18, 2021; date of current version February 5, 2021. This work was supported in part by the Swedish Governmental Agency for Innovation Systems and in part by the project 19ENG02 (Metrology for future energy transmission) from the EMPIR program co-financed by the Participating States and from the European Union's Horizon 2020 research and innovation program. The Associate Editor coordinating the review process was Ron Goldfarb. appropriate antennas; pressure waves can be recorded using acoustic sensors; generated heat can be picked up by thermal imaging; and so on. One measurement technique may not necessarily detect all forms of PD-some sensors are more effective than others in observing specific PD phenomena. All of these different techniques are correlated with the PD phenomena but are impacted by a range of influential variables, including the coupling device (sensor response), measurement circuit (acquisition/sampling unit, data transfer components), signal postprocessing and filtering, and prevailing ambient conditions during measurements, just to name a few. Ideally, a system is completely PD free. However, some degree of imperfections always exists, which may eventually trigger PD events. Unfortunately, outside controlled laboratory test conditions, the onset, the source, the location, and, most importantly, the severity of PD are unknown in advance. In worst case scenarios, improper measurement techniques may be blind to PD events, making a system appear to be PD free. Even, upon detection of PD, correlating the registered data to the severity of the event is challenging, i.e., what is considered acceptable levels of PD. This article is an extension of the CPEM 2020 proceedings article [1] investigating challenges associated with measuring PD using wideband high-frequency current transformers (HFCTs). Such sensors are simple in design and can be manufactured with relatively minimal resources. Installation and utilization are also straightforward as split-core solutions allow for clamping around conductors without the need for service interruptions, disassembly, and tools. For PD applications, in addition to the ability to detect the small rapid event, key functions for a sensor also include the ability to provide information on the defect type and estimate its severity using derived parameters from the recorded signal. This article demonstrates the validity and empirical limitations of existing conventional practices and techniques extending beyond standard specifications to help shape the direction for the future development of wideband PD sensors.
II. PD MEASUREMENT PRINCIPLES Conventional ac measurement techniques for PD (Phase Resolved PD Analysis, PRPDA) are well-established and described in IEC 60270 [2]. Fundamental quantities related to magnitude and phase angle can be derived from measurements and analyzed for fault discrimination and fingerprinting. The PD current pulse i (t), expressed as a double-exponential function, provides charge q as IEC 60270 defines ranges for wideband instruments with lower frequency limits 30 kHz ≤ f 1 ≤ 100 kHz, upper limits f 2 ≤ 1 MHz, and bandwidth 100 kHz ≤ f ≤ 900 kHz. Such wideband instruments observe well-damped oscillatory responses enabling the determination of apparent charge and the polarity of the pulse. Pulse resolution (shortest time interval between two consecutive pulses) is typically 5-20 μs. Narrowband instruments (bandwidth: 9 kHz ≤ f ≤ 30 kHz and midband frequency: 50 kHz ≤ f m ≤ 1 MHz) produce oscillatory responses with envelope values proportional to apparent charge but independent of the polarity. Pulse resolution is typically longer than 80 μs.
Despite varying instrument responses, within the confines of IEC 60270, the concept of "apparent charge" and "calibration" is valid. Calibration is performed to determine the scale factor k for the measurement of the apparent charge by injecting rapid current pulses of known charge magnitude into the terminals of the test object. A PD measuring system working in this "low" frequency range detects the constant part of the PD pulse frequency response. "Since the upper cutoff frequency of the detection bandpass is significantly lower than the upper cutoff frequency of the pulse frequency response, the detected PD pulses are directly proportional to the apparent charge of the PD current pulse" (quasi-integration) [3]. When the test object can no longer be represented as a simple lumped capacitance, and frequencies extend into the HF-UHF regime, the validity of calibration (correlation to apparent charge) becomes void. This occurs for larger complex devices, such as transformers, rotating machines, long cables, and gas insulated switchgear (GIS), which exhibits transmission line characteristics [4]. Here, only a portion of the PD energy arrives at the sensors, whereupon the PD magnitude becomes a relative measure of PD activity and the "calibration" is replaced by "normalization." A direct relation between voltage (mV) and charge (nC) is questionable as the observed signals originating from an unknown location are distorted by pulse propagation phenomena (reflections, resonance, cross-coupling, and so on) [3].
A higher low-frequency cutoff threshold increases the probability of neglecting PD signals distant from the sensor since high-frequency components of the signal are considerably attenuated and may become completely undetectable [3]. Nevertheless, the push for wideband operation extending to higher upper frequency ranges is motivated by interference rejection and suppression principles. Lower frequency ranges are subject to more noise and disturbances (a larger scope of signal sources, including undesired signals) [5]. This is particularly relevant when the voltage supply itself produces rapid excitations, e.g., converter fed motor switching with repetitive short rise time voltage impulses. Furthermore, dc applications, which lack phase information enabling conventional PD quantity derivation, implement pulse shape analysis based on, e.g., rise time, decay time, and pulsewidth [3]. Waveform parameters derived from individual PD pulses can be used for the characterization of different defects using clustering techniques for the separation of multiple sources (defects) and noise [6]- [12]. Pulse shape analysis, thus, requires highresolution details and a clear understanding of the sensor characteristics with which said signals are observed.

III. HFCT DESIGN
Current transformers are inductive couplers widely used as nonintrusive measurement devices converting time-varying current into a voltage (or current) signal scaled by the number of turns in the transformer. Power-frequency applications generally implement grain-oriented silicon-iron cores, while HFCTs utilize metallic oxide materials (ferrites) [13]. HFCTs are often installed on the ground connection of a device where a current flowing along the conductor through the HFCT (single turn primary) induces a voltage measured across a resistive load. This induced voltage in the secondary is proportional to the rate of change of current in the primary [15]. Although simple in operational principle, design characteristics play a crucial role in the suitability of the HFCT for partial discharge applications. According to [17] and [26], sensors with flat and wideband frequency responses allow for the detection of pulse shape-related features but have a tradeoff in reduced gain. For circumstances where accurate pulse shape is not a priority, sensitivity can be improved by designing a higher gain HFCT sensor with a nonflat ("peaky") frequency response. Here, one must acknowledge that such a sensor distorts the pulse shape since different frequency components are amplified and phaseshifted by different levels.

A. Core Material and Winding Design
Soft ferrite cores, such as manganese-zinc (MnZn) and nickel-zinc (NiZn), are effective couplers between electric current and magnetic flux [14]. The sensitivity of the sensor is significantly improved using such materials. However, ferromagnetic cores introduce nonlinearity to the transfer function (dependent on e.g., frequency, temperature, and flux density) [15]. Initial investigations assess variability within the N30-type MnZn cores shown in Table I. Relatively similar physical dimensions were selected to allow for mounting within the same sized enclosures.  Winding type and the number of turns were varied for "HFCT A" designs-copper adhesive tape of widths 6, 12, and 25 mm was compared with 0.22-mm 2 covered conductor winding. HFCT applications typically implement a low number of turns to extend the effective operating range (upper cutoff frequency) to higher values. Since primary turns are generally fixed as N p = 1 (ground conductor), increasing the number of secondary turns N s shifts operation to lower frequencies and reduces sensitivity, as apparent in Fig. 2. Significant variation between wire and strip (tape) windings was not observed. Based on the flatness of the transfer impedance for N s = 5, this design was selected for further development.

B. Shielding-Aperture Design
An aperture (slit in the sensor enclosure) is said to improve coupling between the current-carrying conductor and the winding [16] and prevents the formation of a short circuit turn around the core. Several designs were investigated (see Fig. 3). For apertures located within the inner diameter of the ferrite core, aperture size influences the sensitivity of the sensor (design B)-larger aperture correlates to higher gain. As expected, the highest sensitivity is achieved without any shielding between the current-carrying conductor and the  enclosure (design A). Interestingly, design C exhibits near identical characteristics to design A as evident by the overlaid plots in Fig. 4. Allowing for circulating currents using twopoint grounding (short-circuit turn around core) is effective in hindering coupling up to 1 MHz, whereupon sensitivity returns to similar values as with the other designs.

C. Bandwidth
Transfer impedance and low-frequency cutoff values were determined by measuring the output voltage of the HFCT (NI PXIe 5164, 400 MHz, 1 GS/s) and dividing by the measured current (Tektronix TDS3054B 500-MHz digitizer; TCP202 50-MHz current probe). The sinusoidal current of varying frequency up to 15 MHz was supplied by an HP 33120A signal generator to a 50-termination via short twisted conductors (one passing through the HFCT and the other for return current). Results for N30-type cores are shown in Fig. 5. HFCT B and HFCT C have slightly lower sensitivity compared to HFCT A. Increasing the number of cores (HFCT B double core) to increase inductance does not result in larger gain but is noticeable as a lower cutoff frequency.
HFCT A was selected as a reference for N30-type cores and compared to materials with varying permeability and inductance (refer to Table II). As evident in Fig. 6, all three sensors exhibit similar gain magnitudes (the same number of turns and the same enclosure design), while core selection is evident in the frequency domain-higher permeability and inductance extend the frequency range to lower values (and vice versa).
The aforementioned technique for defining transfer impedance is limited to 15 MHz by the signal generator. To obtain upper frequency values, an Agilent Technologies E5061B, 5 Hz-3 GHz, vector network analyzer (VNA) was utilized to measure S-parameters for HFCT A, D, and E  (see Fig. 7). Low-frequency limitations of the VNA electronic calibrator module (Agilent N4690-60004, 300 kHz-18 GHz), along with test assembly grounding contacts requiring disassembly and reconnection for mounting each HFCT, caused instability in the measured data, which was observed as a fluctuating offset or overshoot of several dB in measured  values (i.e., gain) in the frequency range of 1 MHz. Nevertheless, lower cutoff frequencies based on measured S 21 parameters coincided well with those obtained from the transfer impedance measurements. S-parameter values above 1 MHz were stable. The influence (attenuation) of the test assembly without the HFCT was also measured and was removed from the recorded S-parameters, which results in slightly reduced upper cutoff frequency values. The frequency-amplitude spectrum of the three HFCTs is presented in Fig. 8, and relevant properties are summarized in Table II. Reliably measuring parasitic capacitance is difficult and, thus, causes a discrepancy in calculated theoretical values for the upper cutoff frequencies compared to measured empirical values.

IV. PD APPLICATIONS
In general, PD impulses are characterized as short rise time narrow pulses with varying frequency spectrums, mostly under tens to hundreds of megahertz, but some extending into the UHF-regime (dependent on defect type, test object, applied voltage stress, measurement and acquisition techniques, and so on) [8], [13], [17]- [25]. Comparability of research data is limited and a product of each measurement setup and associated transfer functions. Even for a fixed test assembly, PD data are stochastic by nature. To enable characterization of the developed HFCT designs, a repeatable stable PD source was needed. An Omicron CAL 542 calibrator (1-100 pC) was used to supply a repetitive, c. 4-ns rise time, pulse directly to a NI PXIe 5164 digitizer via short conductors, as illustrated in Fig. 9.
For 100-pC calibrator pulses, HFCT A consistently observed slightly slower waveforms, while HFCT D and E responded with very similar characteristics (see Fig. 10). The recorded calibrator signal's highest significant frequency component is at 37-41 MHz, within the bandwidths of the HFCTs. All sensors detect rather similar peak values, with  the ratio between calibrator peak value and HFCT values, on average, 6.3 (standard deviation s = 4.1%), 6.1 (s = 4.3%), and 6.3 (s = 2.0%) within the entire 1-100 pC calibrator range for HFCT A, D, and E, respectively. The signal-to-noise ratio (SNR) worsens for the smaller calibrator injections 1-5 pC with zero-level and front oscillations hindering the accurate determination of peak values and rise time, as demonstrated in Fig. 11, for 1-pC charge injections.

A. Impact of Amplification
Injecting such calibrator pulses into circuits, including resistive, capacitive, and inductive components, external interference sources, and other "real-world" nonidealities, further alters the characteristics of the signal and can mask relevant features in excessive background noise. In addition to interference mitigation techniques, amplification may be needed to meet the most stringent acceptance criteria.
A 1-pC calibration pulse was injected across a 0.5-pF dummy test object in a conventional PD test setup. The HFCT signal was fed through a Tektronix AM502 differential amplifier (10-kHz-1-MHz bandwidth) to the NI PXIe 5164 digitizer. The amplifier gain was varied to determine the linearity of the amplification range. As evident in Fig. 12, the slower response of the amplifier considerably alters the HFCT response. Nevertheless, waveform characteristics of the amplified signal remain constant for 100-5000 gain factors. The amplification of zero level noise for the highest gain settings results in some distortions, but peak detection of the amplified signal is considerably improved. However, the slower response has an impact on achievable pulse repetition rates. The amplified signal begins to distort the pulse train response at 250 kHz, while the unamplified HFCT signal is still distinguishable above 2 MHz repetition rates (see Fig. 13).
Such superposition of subsequent pulses can result in misleading PD analysis, as demonstrated in Fig. 14, when Influence of differential amplifier on 1-pC calibrator injection measured by HFCT. measuring Trichel pulses (negative corona). As applied ac-voltage is increased, the occurrence of Trichel pulses increases--greater number of discharges within the negative half-cycle visualized as a wider area of consistent amplitude pulses. The accumulation of signals at the extremities of the pulse distribution for C2 and C3 is not a physical phenomenon related to the PD process, but instead a consequence of signal superposition due to insufficient settling time between discharge pulses.
The pulse repetition rate has a significant impact on a key PD parameter--i.e., charge estimation. Mor et al. [26] apply two approaches for estimating charge, a "simplified HFCT model" and a "generic HFCT model" where both techniques implement voltage proportionality to total charge in relation to integration time. The simplified model is given as The gain H is calculated as RM/L 2 , where R is the loading resistance and L 2 is the secondary winding inductance. The mutual inductance M is calculated from the linear slope of the HFCT transfer function in the low-frequency range. For Signal distortion caused by slow sensor response to repetitive discharge. C1 is applied voltage, slower C2 and C3 signals are obtained with IEC 60270 compliant coupling devices, and C4 is a shunt prototype with a fast response but poor sensitivity. Top: initial onset of Trichel pulses. Bottom: misrepresentation of phase-resolved PD data due to poor pulse resolution of coupling devices. the generic model The integration time (approximated as 4L 2 /R) is related to the poles of the sensor transfer function, which defines lower and upper cutoff frequencies ( f lower = R/2π L 2 and f upper = 1/2πRC). Flat wideband sensors require longer integration times compared to those with a peaky response. Fig. 13 demonstrated that sufficient deadtime is achieved to discriminate between subsequent pulse shapes. However, for charge estimation, pulse repletion rates faster than the integration time will include multiple signals and, thereby, result in erroneous values. Shorter integration times are achieved by shifting the lower cutoff frequency (proportional to R/L 2 ) to higher values. This can be achieved with smaller L 2 by reducing the number of turns or alternatively selecting a core with a smaller inductance (L 2 = LN 2 ). Although a smaller core inductance L corresponds to a larger gain, the sensor bandwidth is decreased. Increasing lower cutoff frequencies may be desirable for filtering disturbances but can, in turn, risk losing relevant PD information related to the event of interest. One of the fundamental design questions is, thus, related to the intended purpose of the sensor-pulse shape analysis or magnitude (charge) estimation.

B. Impact of HFCT Placement
Despite being nonintrusive and nonloading, the installation location of an HFCT impacts the observation of PD events. The influence of placement and grounding was investigated by varying configurations within the conventional IEC60270 Fig. 15. Conventional PD circuit with varying HFCT location along the ground conductor of dummy test object C a (parallel disk-electrode gap, c. 0.5 pF). The presence (or absence) of coupling capacitor C k does not influence the performance of the HFCT. circuit shown in Fig. 15. From the perspective of waveform integrity (pulse shape analysis), the dominant characteristic for HFCT applications intended for wideband PD pulse acquisition appears to be related to traveling wave propagation. In a typical context, one often considers the transmission characteristics of traveling waves as they propagate from the source to the measurement point, e.g., time-differenceof-arrival (TDOA) analysis for underground cables, overhead lines, or GIS. For HFCT installations on the ground path of circuits, it is necessary to consider what happens to the signal after it passes the sensor location. That is, in addition to transmission within the test object, the ground conductor and, in particular, the ground termination need to also be considered part of the transmission path. Good grounding practices strive to minimize ground conductor lengths. Results suggest that the length of the ground conductor between the test object and laboratory ground common point is not as critical as the position of the HFCT along said length of ground conductor (for PD applications).
HFCT D was placed immediately after the test object C a , while HFCT E was displaced at intervals toward the common ground point 5.5 m from the test object. The test circuit was situated in a high-voltage laboratory with "good" (low resistance) grounding. Nevertheless, for the frequency spectrum of the emitted signals, the laboratory ground appears to be an open connection or a high impedance. This impedance mismatch results in a discontinuity region at the ground interface, which is seen as a near full reflection in Fig. 16(c) for HFCT E situated close to the ground common point. Time delay in the HFCT E signal is also consistent with propagation velocity along the bare ground conductor length (5.5 m at 3· 10 8 m/s gives 0.18· 10 −7 s). The midway point [see Fig. 16(b)] shows a double peak caused by the superposition of the emitted signal from the test object and the reflected signal from the ground point. Almost identical signals are measured at the same location immediately after the test object C a [see Fig. 16(a)].
The same circuit was used to evaluate the applicability of the charge estimation techniques described in [26]. As expected for oscillatory signals, the resulting integrals are also oscillatory. The charge is estimated from the integral's steady-state  value, but this is not clearly distinguishable from the measured data. Single-integral values for charge shown in Table III are calculated from a restricted time period where oscillations at the peak of the integral give weakly steady-state average values prior to the onset of decay. Setting integration limits (t 1 , t 2 ) manually for each measured signal is not practical as the response changes based on the sensor's interaction with the circuit and test object and is open for interpretation. Methods to mitigate oscillations by means of low-pass filtering to facilitate this approach are discussed in [27]. Without any filtering or processing, HFCT D, which is placed in the immediate vicinity of the test object C a , approximates the charge as roughly 92 pC using the double-integral approach for a 100-pC injected pulse, as shown in Fig. 17.
However, HFCT E, which is displaced along the length of the ground wire, is unable to converge to a value   approximating the target charge of 100 pC. As evident in Fig. 18, the single-integral signal has returned to zero-level signifying peaking of the double-integral for distances 3 and 5.5 m. Distance 0 m would require a longer integration time for HFCT E to determine a peak value using the double-integral approach. Despite not providing a reliable charge estimation, the impact of sensor location is evident in Fig. 18 where charge calculations based on the peak of the double-integral vary depending on the placement of the HFCT sensor along the ground wire. This once again emphasizes that not only is the traveling wave phenomenon relevant within the test object itself, something as seemingly insignificant as placement of an HFCT sensor on a relatively short length of ground wire outside the test object contributes to the accuracy of PD assessment.
Neglecting the impact of the measurement circuit and traveling waves (measurement according to Fig. 9) does not improve the validity of the charge approximation methods for the  Table IV). Fig. 19 shows charge approximations for 100-pC pulses measured by the HFCT sensors. Required integration times for the double-integral method using the approximation t = 4L 2 /R far exceeds the measurement duration. Although not completely accurate (HFCT D peaks at 2.4 µs instead of the estimated 8.84 µs in Fig. 17), the approximation is indicative of the differences between the integration times of the constructed sensors as a function of their physical characteristics (inductance L) and can be used for design purposes. Longer acquisition times can extend evaluation periods, but, as described earlier, this comes at the expense of pulse repetition rates. Furthermore, the compromise between record length and the sampling rate is a considerable challenge and hardware restriction for performing pulse shape analysis together with charge estimation of PD events.

V. CONCLUSION
Despite the vast innovative work performed by all in this field, PD still largely remains a case-specific event as demonstrated in this article. In particular, the concept of measurable "true" charge is illusive, and the validity distinction between calibration and normalization is complicated. Continued efforts are needed for better focused measurement procedures and sensor designs to enable unified practices and universal applications.
It is clear that high-frequency signals attenuate rapidly with distance, and therefore, sensors need to be close to the PD activity for adequate sensitivity. However, larger peak values can be observed farther away from the PD source, as demonstrated by the constructive superposition of signals at the ground termination. This is not correlated with the physical event but rather the local site characteristics and grounding practices.
The sensitivity of the constructed HFCTs without amplification is sufficient for detecting PD signals tens of mV in magnitude injected into a controlled laboratory test circuit with a low-capacitance test object. The sensitivity needs to be improved for better low-level PD detection in more realistic application environments. Increasing gain by altering physical characteristics, such as winding turns and core materials, can migrate sensor characteristics away from the flat response desired for pulse shape analysis. For charge estimation, avoiding summation of multiple subsequent signals requires that integration times do not overlap with pulse repetition rates. However, interdependencies limit the design of sensors optimized for both accurate charge estimation and representative pulse shape analysis. The fundamental design question for PD applications using HFCTs is, thus, related to the intended primary purpose of the sensor. One may strive to combine both conventional coupling devices for charge estimation and wideband sensors for pulse shape assessment into the same circuit, but conventional devices load the circuit and, thereby, alter the waveform, whereupon HFCT placement becomes critical. If using multiple HFCT sensors distributed within the circuit, placement along grounding paths shall be similar (symmetrical) to allow for comparison.