Online Monitoring of Partial Discharges in Large Power Transformers Using Ultra-High Frequency and Acoustic Emission Methods: Case Studies

Abstract

Partial discharges (PDs) are one of the leading causes of catastrophic power transformer failures. To prevent such failures, online PD monitoring systems are increasingly being implemented. In this paper, to the best of the authors’ knowledge, a case study analysis of short-term PD monitoring is presented for the first time using a combination of acoustic emission and ultra-high-frequency methods. Studies have shown that this approach, supported by selected statistical methods for analyzing the convergence (such as the confusion matrix and agreement metrics) of acoustic and electromagnetic pulse detection, improves the reliability of PD detection. Furthermore, it was shown that short-term PD monitoring enables the identification of time windows during which discharges occur periodically and the determination of the transformer phase containing the PD source. This, in turn, facilitates the application of the time difference of arrival (TDoA) technique for the precise localization of transformer insulation defects.

1. Introduction

The reliability of large power transformers is a key element in ensuring stable electricity supplies to consumers. A significant problem that affects many countries around the world is the high percentage of units that have exceeded their designed service life, usually estimated at around 30–35 years [1]. Due to the high costs of purchase, transport, and insurance, as well as the long order lead time, power transformers are not immediately replaced with new ones after their service life is exceeded. The high costs of replacing the device encourage the operators of the power system to extend its service life as much as possible, unfortunately resulting in an increased risk of failure [2,3].

According to a report prepared by the CIGRE association, PDs are one of the most common insulation degradation mechanisms leading to transformer failures [4]. The dynamics of the development of a defect caused by PDs depends on factors such as the type of discharge and its energy, the time of the impact of the discharge on the insulation system, the degree of aging and moisture of the insulation system, the oil temperature, and the moisture migration process [5,6]. In some cases, the defect may develop over many years of transformer operation without leading to its failure. In other cases, the defect may develop extremely quickly, most often as a result of random events, such as overvoltages or a sudden increase in oil temperature caused by unit overload or cooling system failure. Often, the dynamics of the development of the defect accelerates in its final stage, just before failure. For this reason, in recent years, online PD monitoring systems have gained importance [7,8,9]. The main advantages that give online PD monitoring an advantage over periodic tests are the following:

• The ability to immediately detect PD ignition and a sudden, sharp increase in their intensity.
• The ability to detect an upward trend in the values of parameters describing PD pulses, thanks to which the maintenance department can quickly develop and implement procedures that minimize the risk of failure.
• The ability to adjust the schedule of periodic diagnostic tests (e.g., accelerating or postponing them) and to plan downtimes and maintenance work based on data recorded by the PD monitoring system.
• The ability to detect the relationship between PD activity and transformer operating parameters, such as voltage, oil temperature, or on-load tap changer position, which in turn opens the possibility of forecasting PD activity based on statistical models and artificial intelligence methods.
• The ability to estimate the approximate location of the PD source based on the analysis of the activity of acoustic emission transducers and then precisely determine the location of the defect using the time difference of arrival (TDoA) technique.

Currently, the most popular direct PD detection methods adapted for online monitoring include the acoustic emission (AE) method [10,11] and the ultra-high-frequency (UHF) electromagnetic method [12,13]. Both methods have their strengths, but also limitations, which is why neither of them can be considered an ideal and sufficient technique to monitor PDs. For this reason, in the aspect of advanced transformer diagnostics, the simultaneous use of both methods is increasingly suggested [14,15,16,17,18,19]. The main advantage of this strategy is primarily the increased reliability of discharge detection. This results from the fact that the methods discussed are based on different physical phenomena and are susceptible to different types of interference. Sources of acoustic noise usually do not generate electromagnetic signals, while sources of electromagnetic interference generally do not generate acoustic signals. During transformer monitoring using the AE method, potential sources of acoustic interference include oil pumps, fans, core (the so-called Barkhausen noise), intensive weather phenomena (e.g., heavy hailstorms), and tap changer operations. In the UHF method, the source of interference may be transmitters of radio stations and civil services, air navigation systems and transmitters of DVB-T digital television operators, and GSM mobile telephony operators. Therefore, if the measuring system records only pulses of one type (UHF or AE), their source may be both interference and PDs. In turn, if the system simultaneously records UHF and AE pulses, with correlated changes in their parameters (such as amplitude, energy, and frequency), it can be assumed with a high probability that these signals originate from PDs. Additionally, the reliability of this diagnosis can be increased by finding a correlation between PD parameters and basic transformer parameters recorded by the superior SCADA system.

In the last decade, research has intensified on the simultaneous use of two or more PD detection methods to improve the reliability of the evaluation of the condition of the transformer insulation system. A large part of the publications concerns the issue of the localization of PD sources using the TDoA technique, which involves the analysis of the differences in the arrival times of acoustic signals to AE sensors in relation to the radio signal recorded by a UHF antenna or a high-frequency current transformer. The research results have shown that the accuracy of the location of the PD source using the combined electromagnetic–acoustic method is higher than that of a fully acoustic system with four or more ultrasonic transducers [20]. A relatively new area of scientific research is related to the development of combined PD detectors. Siegel et al. proposed a detector consisting of an antenna and an AE sensor, which is introduced into the transformer tank through an oil valve [21]. A similar design of the in-oil PD sensor was also presented by Wenrong Si et al. [22]. In turn, Xiaohu et al. proposed a sensor that comprises a conformal L-shaped ultrasonic array with 13 elements and an ultra-high-frequency electromagnetic sensor [23]. This L-shaped array creates 97 virtual elements through advanced high-order cumulant processing technology, which improves the aperture and direction sharpness of the array. In [24], Sikorski presented an active dielectric window (ADW), which is a new concept of a combined acoustic emission (AE) and an electromagnetic PD detector intended for assembly in a transformer inspection hatch. The ADW has been optimized for operation in mineral oil, which ensures higher PD detection sensitivity than the universal AE sensor mounted outside the transformer tank. In [25], Firuzi et al. presented a novel approach to monitoring power transformers using PD measurements simultaneously according to the IEC 60270 standard [26] and radio techniques. The authors propose a method that allows one to distinguish between internal sources of discharges and external disturbances, as well as to identify different internal sources of PD.

Online PD monitoring using the AE and UHF methods simultaneously is such a new issue that there is a lack of publications that present in detail the practical experience collected in this area. In the work of [27] published in 2020, the authors presented the first online PD monitoring system, whose operation was based on three diagnostic methods (HF, UHF, and AE). The prototype of the multi-sensor monitoring system was first validated on a large 330 MVA power transformer during an induced voltage test with partial discharge (IVPD) measurement. Then, it was mounted on a 31.5 MVA substation power transformer and integrated with the SCADA system that records the voltage, active power, and oil temperature of the monitored device. The results obtained confirmed the advantages of the simultaneous use of three PD detection techniques and the possibility of correlating discharge parameters with other parameters of the power transformer. In 2021, Ranninger and Krüger in [28] presented the results of a 3-month monitoring period of a 300 MVA transformer, as a result of which it was possible to locate a defect near phase 1U, which was the PD source. In 2022, Gorgan et al. in [29] discussed the simultaneous application of conventional and alternative PD detection methods in power equipment diagnostics. The authors compared the sensitivity of each method in the context of detecting critical insulation defects. In 2024, Mehrjou and Hosseini in [30] presented the idea of a combined method for partial discharge detection and localization in power transformers using only one acoustic emission and UHF sensor. Simulated in COMSOL Multiphysics 5.2a and experimentally validated, this approach uses pattern recognition in MATLAB to compare real-time sensor data with a predefined database, ensuring accurate PD detection and localization.

The significant contribution of this work is, to the best of the authors’ knowledge, the first detailed analysis of several cases of online PD monitoring in large power transformers using both ultra-high-frequency and acoustic emission methods. Another contribution is the presentation of fault location results obtained using the TDoA technique and based on the analysis of data collected by the acoustic emission system, which allowed the search for the PD source to be narrowed to a small area of the transformer tank and significantly reduced the measurement time.

2. Materials and Methods

2.1. Online Monitoring of PD Using the UHF Method

2.1.1. Detection of PD Using the UHF Method

The physical basis of the UHF method is the detection of electromagnetic field disturbances caused by a PD current pulse with a short rise time (<1 ns). Microstrip and printed antennas, monopoles, aperture antennas, and disk couplers are mainly used to detect electromagnetic radio waves propagating inside the transformer tank. The sensitivity of PD detection using the UHF method is assessed as high, but it largely depends on the type and design of the antenna used, the distance between the antenna and the PD source, the method of mounting the antenna in the transformer tank (in a special dielectric window or inside the oil valve), and the presence of metal obstacles between the antenna and the PD source. The main advantage of the UHF method is its greater resistance to electromagnetic interference than the conventional IEC 60270 method, as most of the energy from electromagnetic noise in the power system is transmitted in the range of relatively low frequencies, below several MHz. In addition, some measurement systems allow for the use of an external, broadband noise-gate antenna to reject interference. Another valuable advantage of the UHF method is the possibility of obtaining phase-resolved partial discharge (PRPD) patterns that are used to identify the type of PD, as well as potential electromagnetic interference. It is worth mentioning that in the case of installing the antenna in an oil valve or in special dielectric windows (which are increasingly often transformer factory equipment), there is no need to turn off the monitored device, which, especially from the perspective of the power system operator, is another advantage of the UHF method. The main disadvantage of the UHF method is the lack of a reliable measurement of the apparent charge of PDs and the high price of the measurement system [12,31,32].

2.1.2. Parameters and Functions of the UHF PD Monitoring System

The Qualitrol Portable PDM system (Qualitrol Company LLC, Fairport, NY, USA) was used for UHF partial discharge monitoring (Figure 1a). The system has four measurement channels, the first three of which are used to connect PD detectors operating in the VHF/UHF band, while the last channel is designed to connect a special antenna that reduces external electromagnetic interference. The dynamic sensitivity range of the measurement channels ranges from −35 to −75 dBm, and the amplitude of the recorded pulses is related to the maximum value of the measurement range and is given in a percentage. The system software records the number of pulses, amplitude, and PRPD pattern of UHF PD pulses, the image of which is used in the defect identification (PD-type) process. To detect UHF pulses, a Qualitrol CMD01591-type wideband rod antenna (Figure 1b) with a bandwidth of 200 to 1500 MHz (Figure 1c) was used, which was placed in the oil drain valve. The measurement system was additionally equipped with a gating antenna, which allowed filtering of unwanted external electromagnetic interference. Implementation of the gating antenna technique allowed for a more effective detection of UHF pulses from PDs by minimizing the influence of irrelevant or transient RF signals.

2.2. Online Monitoring of PD Using the AE Method

2.2.1. Detection of PD Using the AE Method

The AE method relies on the detection and measurement of impulse pressure changes generated during PD events in the insulation system. Acoustic emission waves are mainly detected using contact piezoelectric transducers, which are installed on the outer wall of the transformer tank using magnetic holders. Some types of AE transducers are designed to be installed inside the transformer tank, which, in theory, is supposed to provide a higher sensitivity to PD detection. These are primarily optoacoustic sensors, active dielectric windows, or hydrophones. The AE method has the following advantages: (i) resistance to electromagnetic interference, (ii) measurements that are non-invasive (installation of piezoelectric AE sensors on the tank does not require switching off the transformer) and relatively easy to perform, and (iii) low price of the measurement system, especially in comparison to the UHF method and the conventional electrical method. The main disadvantage of the AE method is the lack of the possibility of calibrating the acoustic signal with respect to IEC 60270. The sensitivity of the AE method is directly influenced not only by the distance of the piezoelectric transducer from the PD source but also by the phenomenon of attenuation, diffraction, interference, and scattering of the acoustic wave. For this reason, PDs of low energy (<300 pC) or located in deep layers of the insulation system are difficult or impossible to detect [33,34].

2.2.2. Parameters and Functions of the PD Monitoring System Using the AE Method

The PDtracker Portable system (Figure 2a), developed by Poznan University of Technology, was used for online PD monitoring using the acoustic emission method. The system enables simultaneous sampling on 8 analog inputs with a maximum resolution of 12 bits and a speed of up to 20 MS/s (millions of samples per second). Piezoelectric transducers of type A6890 (Poznan University of Technology, Poland) (Figure 2b) were used to record AE pulses, which are optimized for the detection of PD in oil–paper insulation. The A6890 transducer has two piezoelectric disks with opposite polarization directions and different heights. This fully differential design allows this transducer to obtain (i) high peak sensitivity (−61.1 dB ref. V/µbar), (ii) two resonant frequencies for longitudinal acoustic waves (68.7 kHz and 89 kHz), which are characteristic of surface discharge and discharges in transformer oil, respectively, (iii) a wide bandwidth from 20 to 100 kHz (Figure 2c), and (iv) low noise (removal of the common-mode noise by the differential preamplifier) [35]. A differential preamplifier with a gain of 40–60 dB and an active band-pass filter (20–500 kHz) is integrated with the piezoelectric transducer in a waterproof housing equipped with four magnetic holders. The PDtracker Portable system software (version 1.6) was developed in the NI LabVIEW 2019 environment. One of the main tasks of the software is to determine, in real time, five parameters of AE pulses from PDs, i.e., the number of pulses per minute, the amplitude and average energy of pulses, and the amplitude and maximum energy of pulses.

2.3. Localization of PD Sources Using the TDoA Technique

For each of the power transformers discussed in this article, in addition to online monitoring, the location of PD sources was determined using the TDoA method. This is one of the most popular and widely used location techniques, the detailed principles of which are presented, among others, in [36,37,38]. The basis of the TDoA method lies in the observation that the acoustic wave generated by PDs propagates inside the oil-filled transformer tank in a manner close to spherical. The wave reaches different transducers with varying time delays, depending on the distance from the PD source to each sensor.

The TDoA localization procedure consists of the following steps: (1) Registration of the time waveforms of acoustic emission signals generated by PDs using at least four piezoelectric transducers mounted on the transformer tank. (2) Estimation of the onset times of AE pulses. (3) Determination of the time differences of arrival of the acoustic wave between the reference transducer (i.e., the sensor located closest to the PD source, which is the first to register the AE pulse) and the other transducers. (4) Based on the known coordinates of the transducers, the time differences of arrival of the signals at each transducer, and the velocity of the acoustic wave propagation in transformer oil, a nonlinear system of equations describing the spheres is formulated. The centers of these spheres correspond to the positions of the transducers, and their radii represent the shortest geometric distances between the transducers and the PD source. The intersection point of these spheres indicates the location of the PD source. (5) Solving the nonlinear system of equations using either an iterative or non-iterative method [36].

To locate PD sources in the monitored transformers, the PDtracker Portable system and a dedicated program developed in the MATLAB environment were used. The onset times of AE pulses were automatically determined using the Akaike Information Criterion (AIC) method [39,40]. The AIC method determines signal onset by identifying the point where the statistical properties of the signal, such as variance or autocorrelation, change significantly, marking the transition from noise to the signal. Nonlinear systems of equations were solved using the fsolve function with its built-in iterative Trust-Region-Dogleg algorithm [41,42]. This algorithm is very efficient as it involves only a single linear solve per iteration to compute the Gauss–Newton step [43]. Moreover, it often proves to be more robust than the Gauss–Newton method combined with a line search.

2.4. Case Study Objects

The article presents the results of online PD monitoring carried out on three network transformers, hereinafter referred to as ‘Transformer A’, ‘Transformer B’, and ‘Transformer C’, the parameters of which are listed in

Table 1. Parameters of the tested power transformers.

Transformer Designation	Year of Manufacture	Rated Power (MVA)	Rated Voltage (kV)	Winding Connections	Cooling Type	Test Period
Transformer A	1991	160	230/120/15.75	YNA0yN0	ONAF	17–24 October 2023
Transformer B	1978/1996 1	160	230/120/15.75	YNA0yN0	ONAF	6–7 November 2023
Transformer C	1980	160	230/120	Y0d11	ONAF	23 May–9 June 2024

¹ The transformer underwent a major renovation in 1996.

. Each of the tested transformers was equipped with a standard DIN 50- or DIN 80-type oil valve, and the installation of the UHF antenna in the valve required the use of an additional adapter (extension). The transformers used ONAF (Oil Natural and Air Forced) air cooling technology. There are only two radiators with fans on the high- and low-voltage sides, which made access to the tank surface effortless during the assembly of the AE transducers. This was particularly helpful during the procedure for locating PD sources, as it requires recording acoustic signals with many sensors in various positional configurations.

All three power transformers were selected as test objects by the power grid operator due to concerning results from dissolved gas analysis, which indicated the presence of PD sources based on elevated levels of hydrogen and acetylene (

Table 2. Concentrations of selected gases dissolved in oil in the tested power transformers.

Dissolved Gas	Measured Gas Concentration (ppm)			Typical Gas Concentration Values (ppm) *
	Transformer A	Transformer B	Transformer C
Hydrogen (H2)	177	291	453	150
Acetylene (C2H2)	79	204	67	55
Carbon dioxide (CO2)	–	3401	3216	3000

* Typical gas concentration thresholds set by the power grid operator to indicate proper transformer operation.

).

2.5. Methodology for Online Monitoring of PDs in Tested Power Transformers

In each of the presented cases of online monitoring of discharges in power transformers, the measurement procedure included the following steps:

Step 1.

Switching off the power transformer.

Step 2.

Installing the antenna (UHF PD detector) in the transformer oil drain valve.

Step 3.

Determining the optimal position of the antenna in the oil drain valve. The depth of the antenna was changed in the range of ±3 cm relative to the external plane of the tank wall to achieve the minimum, close to unity value of the VSWR (Voltage Standing Wave Ratio). This procedure aimed to adapt the antenna to optimal operating conditions and minimize signal reflection losses. A portable, vector network analyzer, MEASALL KC901S+ (Measall, Chengdu, China), was used to measure the VSWR.

Step 4.

Installing the noise-gating antenna near the transformer tank.

Step 5.

Installing the eight AE sensors (marked CH1–CH8) on the transformer tank as follows: CH1—on the side wall of the tank, opposite the on-load tap changer; CH2, CH3, and CH4—on the high-voltage side, opposite phases L1, L2, and L3, respectively; CH5, CH6, and CH7—on the low-voltage side, opposite phases L1, L2, and L3, respectively; and CH8—on the side wall of the tank (Figure 3).

Step 6.

Verifying the correct operation of individual measurement channels in the acoustic emission (AE) system and proper acoustic and mechanical coupling of AE sensors using the Hsu–Nielsen method (pencil-lead breaks).

Step 7.

Determining the level of electromagnetic interference recorded by the noise-gating antenna and the UHF PD detector on the de-energized power transformer.

Step 8.

Energizing the power transformer.

Step 9.

Determining the average effective value of the acoustic background noise (A_{noise_RMS}) recorded by the AE sensors on the energized transformer and the threshold above which the monitoring system will count AE pulses, according to the relationship threshold = k · A_{noise_RMS}, where k = 5.

Step 10.

Determining the level of electromagnetic interference recorded by the noise-gating antenna and UHF PD detector on the energized transformer.

Step 11.

Leaving the online PD monitoring systems in operation for the period specified by the transformer owner (from several days to several weeks), during which parameters of acoustic signals (number of AE pulses per minute) and radio signals (number and amplitude of UHF and PRPD patterns) were recorded.

Step 12.

Localizing PD sources using the TDoA technique in the transformer tank area where the AE sensors of the PDtracker Portable system recorded acoustic emission pulses (if PDs were detected when the monitoring systems were activated or during their routine inspection on the following day, acoustic emission monitoring was immediately halted, and efforts were made to locate the PD source).

When two different detection methods are used simultaneously in the study of PD monitoring in transformers, an important issue is to assess the degree of convergence of the results obtained. To perform an objective analysis, it was decided to use the confusion matrix and appropriate agreement metrics, which allow for an accurate determination of how closely the results of both methods agree.

The confusion matrix is a tool that allows for a detailed presentation of the results of two monitoring methods, in this case, UHF and AE, in the form of a table comparing the results of both techniques in terms of their ability to detect PD pulses. In the classic confusion matrix for two methods, where each of them can indicate “1” (PD pulse detected) or “0” (no PD pulse), four basic categories are distinguished:

• True positive (TP): The number of cases in which both methods detected a PD pulse during the same period. This means that both systems correctly recorded the presence of a discharge. In this case, there is a low probability that both the UHF and AE methods were recording interference at the time.
• False positive (FP): The number of cases in which one method recorded a PD pulse and the other did not. This suggests that one of these methods was more sensitive in detecting PD or may have captured interference signals.
• False negative (FN): The number of cases in which one method did not detect a pulse recorded by the other method. Here, the reasons may be identical to those in false positive cases.
• True negative (TN): The number of cases in which both methods did not record any PD pulse. This means that both methods are consistent in identifying periods without discharges or are too sensitive to detect them, e.g., in a situation where the generated PD pulses had low energy.

The confusion matrix in this context enables the assessment of the agreement of both PD monitoring methods, allows the examination of how often both methods agree in their results, and also identifies cases where there are differences in the detection of PD phenomena. On the basis of the confusion matrix, various agreement metrics can be calculated, which allow a more precise determination of the degree of similarity between the results obtained from two measurement methods. Among the various agreement metrics, Accuracy Index and Jaccard Index were selected as they best correspond to the specificity of comparing two simultaneously applied methods for PD detection.

Accuracy Index is one of the basic measures that determines how often both methods give the same results, i.e., both simultaneously record or do not record PD pulses. It is the ratio of the number of cases in which both methods give the same result (TP + TN) to the total number of cases (TP + TN + FP + FN):

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y} \mathrm{I}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{E}\mathrm{N}}}$$

(1)

The Jaccard Index measures the similarity between two data sets, in this case, the results of two monitoring methods. It is the ratio of the number of cases in which both methods detect a PD pulse (TP) to the number of cases in which at least one of the methods detects it (TP + FP + FN):

$$\mathrm{J}\mathrm{a}\mathrm{c}\mathrm{c}\mathrm{a}\mathrm{r}\mathrm{d} \mathrm{I}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(2)

To determine the confusion matrix, Accuracy Index, and Jaccard Index, it was necessary to preprocess the measurement data appropriately. In the first step, the number of recorded acoustic emission pulses from the eight channels was summed. Then, due to differences in the temporal resolution of the measurement systems, where the AE system counts pulses in 1 min intervals, while the UHF system counts them in 5 min intervals, a cumulative aggregation of AE pulses in 5 min intervals was performed. In the final step, the recorded AE and UHF pulse counts were binarized, assigning a value of 1 when pulses were detected and 0 when no pulses were recorded.

3. Case Study Analysis of Online PD Monitoring in Power Transformers

3.1. Transformer A

3.1.1. Results of Online Monitoring of Partial Discharges

Figure 4 presents the results of UHF signal and acoustic emission (AE) recordings obtained simultaneously by eight AE transducers mounted on the tank of the tested transformer. Both systems were activated on 17 October 2023, at 3:00 p.m., and from the moment of activation, a high intensity of PD was observed. In the case of PD testing using the AE method, the maximum number of pulses recorded by the CH1 transducer, positioned near the on-load tap changer, was 4660 pulses per minute, representing approximately 60% of the measurement channel capacity. During the first two days of operation of the monitoring system, moderate intensity AE pulses were also recorded by the CH5 transducer, installed near the L1 phase on the low-voltage side. The remaining AE transducers either did not register any pulses or recorded pulses of low intensity. In the case of PD testing using the UHF method, the system recorded numerous pulses with very high amplitudes, reaching 90% of the maximum value of the measurement range. Periods of high and moderate AE and UHF pulse activity predominated in the monitored transformer, while instances of complete signal loss were observed extremely rarely. It is noteworthy that the periods of highest AE and UHF pulse activity largely overlapped, suggesting that their common source was a defect in the insulation system generating PDs.

As mentioned earlier, PD activity depends on various factors, such as voltage fluctuations, including sudden step increases caused by atmospheric and switching overvoltages, the moisture level in the insulation system, and bidirectional moisture migration between the cellulose insulation and oil, induced by temperature changes (water accumulation at the oil–cellulose interface can lead to the initiation of surface discharges). Additionally, the temperature of the insulation system is influenced by factors such as the transformer’s variable load, ambient temperature, and the efficiency of the cooling system. As a result of these factors, PDs exhibit characteristics of a stochastic phenomenon, making them difficult to predict. Therefore, in the authors’ view, if the transformer testing program includes the localization of discharge sources, this should be carried out immediately upon detection of high and stable AE and UHF pulse activity by the monitoring system. Such a situation occurred in the analyzed case of ‘Transformer A’. On the second day of monitoring, from the early morning hours, a stable, high level of UHF and AE pulses was observed. As mentioned previously, the CH1 transducer, installed near the on-load tap changer, recorded the highest acoustic emission activity.

The PRPD images (Figure 5) recorded during this period exhibited features indicative of PDs, including a clear correlation with the supply voltage phase and a repetitive pattern of pulse amplitude and count within specific phase intervals (0–90° and 180–270°). Importantly, the Qualitrol PDM system software also identified the PRPD pattern as originating from PDs.

Based on these data, at 10:30, a decision was made to discontinue monitoring using the acoustic emission method and initiate PD source localization, which continued until 12:30. Throughout the fault localization procedure, the UHF monitoring system continued to record high PD activity.

For further statistical analysis to compare the performance of both diagnostic methods in ‘Transformer A’, the data presented in Figure 4 (number of UHF and AE pulses) were converted to a binary form (pulse detected = 1; no pulses = 0).

The confusion matrix (

Table 3. Confusion matrix for pulses recorded by UHF and AE methods in ‘Transformer A’.

	AE: Pulse (1)	AE: No Pulse (0)
UHF: Pulse (1)	TP = 59.2%	FP = 16.4%
UHF: No Pulse (0)	FN = 17.9%	TN = 6.4%

) was then calculated, and the selected agreement metrics, that is, the Accuracy Index and the Jaccard Index, were determined. The confusion matrix indicates that both methods simultaneously detected PD pulses in 59.2% of cases (TP). At the same time, 16.4% of cases were observed in which the UHF method detected pulses (FP) while the AE method did not. In 17.9% of the cases, the opposite was true, i.e., the pulses were detected only by the acoustic emission method. In 6.4% of cases, both methods simultaneously detected no pulses (TN), which may indicate either a lack of PD activity or a limitation in the sensitivity of both systems. The Accuracy Index (0.656) suggests that the two methods agreed in about 65.6% of cases, a moderately high convergence of results. The Jaccard Index (0.633) shows that the agreement between the results of the two methods, taking into account cases of detection by at least one of them, is 63.3%. These results suggest that UHF and AE methods are largely complementary, but their performance in simultaneously detecting PD pulses may be limited by differences in detection mechanisms or their susceptibility to interference.

3.1.2. Results of Partial Discharge Source Localization

Based on information obtained from the AE monitoring, it was determined that the PD source was most likely located near the on-load tap changer. Therefore, the localization procedure was confined to this area (the side wall of the transformer tank). To pinpoint the PD source using the TDoA technique, pulses were recorded using four AE transducers in three different configurations of their positions, referred to as Test 1, Test 2, and Test 3. A common Cartesian coordinate system, shown in Figure 6, was adopted for all tests.

The coordinates describing the positions of the acoustic emission AE transducers for each measurement test are provided in

Table 4. The coordinates and designations of the AE transducers used in the individual measurement tests (i.e., Test 1, Test 2, and Test 3).

Test Designation	AE Sensor	AE Sensor Coordinates
		X (m)	Y (m)	Z (m)
Test 1	CH1	1.130	0	0.985
	CH2	1.580	0	0.700
	CH3	1.140	0	0.710
	CH4	1.630	0	0.980
Test 2	CH1	1.130	0	0.985
	CH2	1.580	0	0.700
	CH3	1.250	0	0.360
	CH4	0.560	0	0.720
Test 3	CH1	1.130	0	0.985
	CH2	0	0.590	0.895
	CH3	1.140	0	0.700
	CH4	0	0.790	0.750

. Figure 7 shows exemplary time waveforms of acoustic emission pulses, with the signal onsets marked by a red line, which were automatically estimated using the AIC method. Based on the recorded time series, the coordinates of the discharge source locations were calculated for each test. The graphical representation of the obtained results is shown in Figure 8, while

Table 5. Ranges and modal values of the X, Y, and Z coordinates for the estimated location of the PD source in the tested ‘Transformer A’.

The Range of Determined Coordinates of the PD Source
X (m)	Y (m)	Y (m)
0.32–0.70	0.28–0.77	0.63–1.03
The Modal Value of the PD Source Coordinates
X (m)	Y (m)	Y (m)
0.65	0.33	0.70

presents the ranges and modal values (most frequently occurring values) of the calculated PD source coordinates.

The tests carried out using the TDoA technique confirmed that the most probable source of PD pulses is a defect occurring inside the on-load tap changer or its immediate vicinity, e.g., in the tap lead insulation. Upon completion of the monitoring and localization of PD, the area corresponding to the range of defect coordinates was marked on the tank of the tested transformer for the operational services of the power grid operator, while the modal value, i.e., the most frequently obtained solution, was indicated with an ‘×’ (Figure 9).

3.2. Transformer B

3.2.1. Results of Online Monitoring of Partial Discharges

For ‘Transformer B’, the transmission system operator implemented PD monitoring following the dissolved gas analysis (DGA) results, which revealed an elevated acetylene concentration of 204 ppm. This finding indicated the presence of high-energy, high-temperature PDs of the sparking type, often described as small, localized arcs. Given the nature of the detected gas and the potential for rapid progression of the defect, a decision was made to conduct a short-term PD monitoring campaign limited to 24 h. The restriction of the monitoring period was primarily motivated by safety considerations and the necessity to mitigate the risks associated with operating a transformer in a pre-failure state. The online PD monitoring systems were activated on 6 November 2023, at precisely 2:02 p.m. An analysis of the measurement results, presented in Figure 10, indicated that the UHF system did not detect any PD pulses until 6:05 a.m. the following day. During this period, the acoustic emission (AE) system recorded two brief intervals of AE pulse activity (from 3:52 p.m. to 4:05 p.m. and from 4:31 p.m. to 4:46 p.m.), both characterized by low intensity (<5%), as well as a single event occurring at 2:22 a.m. the next day.

The acoustic emission pulses were primarily registered by transducers located on the low-voltage side (CH5, CH6, and CH7) and near the tap changer (CH1). The CH3 transducer continuously recorded sporadic pulses with an intensity below 0.8% throughout the monitoring period, probably due to setting the pulse counting threshold too low. This transducer was located near the radiator and fans, which are potential sources of acoustic interference. Starting at 6:05 a.m. on November 7, both the UHF and AE systems began detecting stable PD pulses, with a sudden and significant increase in their intensity and amplitude observed at 10:35 a.m. The PRPD patterns for UHF pulses, as recorded by the Qualitrol PDM system, were automatically classified as PDs. As illustrated in Figure 11, the concentration of pulses in the first (0–90°) and third (180–270°) quadrants of the supply voltage cycle is characteristic of surface and creeping discharge or certain types of void discharge within gas inclusions [44].

Based on these observations, it was determined to discontinue PD monitoring via the acoustic emission method and proceed with the localization of PD sources.

To summarize the results obtained by both diagnostic methods, the confusion matrix (

Table 6. Confusion matrix for pulses recorded by UHF and AE methods in ‘Transformer B’.

	AE: Pulse (1)	AE: No Pulse (0)
UHF: Pulse (1)	TP = 5.9%	FP = 9.6%
UHF: No Pulse (0)	FN = 17.6%	TN = 66.9%

) was calculated. The values obtained show that both methods simultaneously detected PD pulses in only 5.9% of the cases (TP), which means low agreement on the scope of simultaneous registration of PD pulses. In 9.6% of the cases, the UHF method indicated the presence of PD pulses (FP), while the AE method did not detect them. In turn, 17.6% of the cases were situations in which only the AE method registered pulses (FN). It should be noted that up to 66.9% of cases (TN) were situations in which both methods did not detect PD pulses.

The high value of the Accuracy Index (0.728) shows that both methods agreed in 72.8% of the cases, indicating a moderately high overall agreement of the performance of these methods. However, the Jaccard Index (0.177) indicates very low agreement in the case of detection of the same PD pulses, considering only the cases in which at least one of the methods registered the pulses. The low Jaccard Index value suggests that the UHF and AE methods for ‘Transformer B’ show significant differences in detecting PD pulses.

3.2.2. Results of Partial Discharge Source Localization

Measurements conducted to locate the defect, which lasted from 10:45 to 14:00, were accompanied by parallel UHF monitoring that revealed very high PD activity. During this period, the average relative amplitude of UHF pulses was 83.9%, and the average number of pulses per 5 min was 998. Since the monitoring system recorded acoustic pulses exclusively near the L3 phase on the high-voltage side, the search for the PD source was confined to this area of the transformer tank. To precisely locate the defect, acoustic emission pulses were recorded at 18 measurement points, designated as P1–P18, whose coordinates are provided in

Table 7. The coordinates of the AE transducers installed at 18 measurement points.

Measurement Point Designation	AE Sensor Coordinates			Measurement Point Designation	AE Sensor Coordinates
	X (m)	Y (m)	X (m)		X (m)	Y (m)	Z (m)
P1	0.09	0	0.73	P10	0.75	0.05	1.00
P2	0.38	0	0.70	P11	0.40	0	0.85
P3	0.41	0	0.91	P12	0.09	0	1.50
P4	0.13	0	0.99	P13	0.11	0	1.51
P5	0.09	0	0.70	P14	0.33	0	0.99
P6	0.10	0	0.86	P15	0.33	0	1.51
P7	0.13	0.01	0.78	P16	0.10	0	1.73
P8	0.59	0.01	1.00	P17	0.33	0	1.70
P9	0.60	0	0.87	P18	0.41	0	1.72

.

The locations of the AE transducers and the coordinate system are illustrated in Figure 12a, while sample time waveforms of PD pulses recorded at points P1, P2, P3, and P4, with the signal onsets marked (red lines), are shown in Figure 12b. Figure 12c presents a graphical representation of the PD source localization results (all solutions to the TDoA equation systems).

Table 8. Ranges and modal values of the X, Y, and Z coordinates for the estimated location of the PD source in the tested ‘Transformer B’.

The Range of Determined Coordinates of the PD Source
X (m)	Y (m)	Z (m)
0.31–0.61	0.08–0.34	0.42–1.32
The Modal Value of the PD Source Coordinates
X (m)	Y (m)	Y (m)
0.44	0.22	0.67

provides the ranges and modal values of the PD source coordinates.

An analysis of the results indicates that the scatter of values in the X and Y axes is small compared to that in the Z axis, which reached as much as 90 cm. This may suggest a widespread defect developing, for example, along the windings, as well as along the pressboard barrier or strip. Given that the Y-axis (depth) values are relatively close to the tank (with the most frequently occurring value being Y = 22 cm), another probable cause of the PDs could be damaged lead insulation (e.g., micro-cracks, material aging, or localized overheating).

3.3. Transformer C

3.3.1. Results of Online Monitoring of Partial Discharges

Figure 13 shows the results of two weeks of monitoring conducted on ‘Transformer C’ from 23 May 2024 to 6 June 2024. Three hours after the transformer was switched into operation in measurement channel CH7 (with the transducer positioned opposite the L2 phase on the lower voltage side), the acoustic emission system suddenly began recording stable pulses of moderate intensity, registering 1079 pulses per minute—approximately 14% of the channel’s capacity. From that moment, that is, from 23 May, 6:23 p.m., to 24 May, 2:15 p.m., the system practically continuously recorded AE pulses of variable intensity (from 0.2 to 21%).

The periods in which the AE system and the UHF system recorded pulses simultaneously overlapped by up to 85.7%, which allows us to assume that their source was PDs. The maximum relative amplitude of the UHF pulses was 45.3% during this period. The next periods in which the CH7 transducer recorded pulses occurred in the intervals from 25 May, 2:55 a.m. to 12:17 p.m., and from 26 May, 2:42 a.m., to 27 May, 4:01 a.m. Activity was also recorded in the CH4 measuring channel (transducer placed opposite the L3 phase on the high-voltage side) in the time interval from 25 May, 3:06 p.m., to 27 May, 2:32 a.m. On 27 May from 1:20 to 1:48 pm, a very high (up to 93%) intensity of AE pulses was recorded in the eight measuring channels. The UHF monitoring system also recorded PD pulses at that time with a maximum amplitude of 79%. One of the possible causes of the sudden ignition of PD could be a switching overvoltage in the transformer field or an atmospheric overvoltage, as numerous lightning discharges occurred during this period, which is confirmed by the map shown in Figure 14.

The PRPD pattern for UHF pulses (Figure 15), recorded at that time, was characteristic of a defect: carbonized tracks at the leads of the winding [38]. After 13:48, acoustic emission pulses were still recorded by the CH4 transducer only, which monitored the L3 phase from the HV side, with the pulse intensity decreasing and not exceeding the level of 39%. The pulses were continuous, undying, and disappeared completely after only 5 days, i.e., on 1 June at approximately 22:00. In the remaining phases on the high-voltage side (L1 and L2), no AE pulses from PD were recorded apart from single events. On the low-voltage side, a high intensity of AE pulses (up to 30%) was detected in phase L2. During the 2 weeks of monitoring system operation, discharges were recorded daily, with the highest activity occurring on 26, 30, and 31 May and on the day of system dismantling (6 June). Therefore, immediately after completing the monitoring, a procedure was carried out to locate the source of partial discharges in phase L2 on the low-voltage side.

The statistical analysis carried out to compare the number of pulses recorded by the UHF and AE methods showed that in as many as 39.7% of cases (FN), only acoustic emission signals were detected (

Table 9. Confusion matrix for pulses recorded by UHF and AE methods in ‘Transformer C’.

	AE: Pulse (1)	AE: No Pulse (0)
UHF: Pulse (1)	TP = 32.7%	FP = 17.3%
UHF: No Pulse (0)	FN = 39.7%	TN = 10.3%

). This could be related to the excitation of piezoelectric sensors during storms and the accompanying heavy rainfall. During the two weeks of the monitoring system operation, there were 4 days with intensive weather phenomena. For comparison, the total time in which only the UHF method recorded pulses was 17.3% (FP). Both methods simultaneously detected pulses in 32.7% of cases (TP), and in 10.3% of cases, their absence was indicated. The Accuracy Index (0.429) shows that the results of both methods were consistent in 42.9% of the analyzed cases, which indicates a relatively low level of overall consistency of the UHF and AE systems. The Jaccard Index (0.364) reflects that only 36.4% of the cases in which at least one method detected a PD pulse include situations in which both methods simultaneously detected the same event.

3.3.2. Results of Partial Discharge Source Localization

The location of the PD sources was determined by searching for the area on the transformer tank (near the L2 phase of the low-voltage side) where the amplitude of the acoustic emission pulses was the highest. The highest ‘loudness’ area was found approximately 60–70 cm below and 10–20 cm to the right of the location where the CH7 transducer was attached to monitor PDs in the L2 phase. Then, we attempted to find a more precise location of the PD source using the TDoA technique. For this purpose, three independent tests were performed. In each of these tests, the transducers were placed in a different location (Figure 16), but always near the previously identified area where the maximum level of acoustic emission signal was recorded. The coordinates of the acoustic emission transducers for each measurement test are listed in

Table 10. The coordinates and designations of the AE transducers used in the individual measurement tests (i.e., Test 1, Test 2, and Test 3) performed for ‘Transformer C’.

Test Designation	AE Sensor	AE Sensor Coordinates
		X (m)	Y (m)	Y (m)
Test 1	CH1	0.860	0	0.265
	CH2	1.195	0	0.330
	CH3	0.990	0	0.480
	CH4	0.695	0	0.490
Test 2	CH1	1.030	0	0.340
	CH2	1.380	0	0.250
	CH3	1.315	0	0.505
	CH4	0.695	0	0.485
Test 3	CH1	0.870	0	0.470
	CH2	0.945	0	0.215
	CH3	1.140	0	0.510
	CH4	1.145	0	0.240

.

During the PD localization procedure, the phenomenon showed stability and high intensity, which allowed for recording a significant number of AE waveform sets in each of the three tests performed: 442 in Test 1, 465 in Test 2, and 122 in Test 3. Example AE pulses with marked onset times are presented in Figure 17, while a graphical representation of the localization results is shown in Figure 18.

The spatial distribution of localization points in all three TDoA localization tests is consistent and shows a high degree of similarity. While the range of solutions for the X and Z coordinates is relatively narrow (16 cm for X and 36 cm for Z), the range for the Y coordinate is significantly broader, reaching 90 cm (

Table 11. Ranges and modal values of the X, Y, and Z coordinates for the estimated location of the PD source in the tested ‘Transformer C’.

The Range of Determined Coordinates of the PD Source
X (m)	Y (m)	Z (m)
0.85–1.01 *	0.05–0.95 *	−0.07–0.29 *
The Modal Value of the PD Source Coordinates
X (m)	Y (m)	Y (m)
0.96	0.76	0.11

* The reported data ranges were determined after excluding outliers.

). These results suggest that the source of the PDs may be in the insulation of the low-voltage winding lead. Figure 19 shows a photograph of the tested transformer, with the area marked to indicate the range of X and Z coordinates of the PD source.

4. Conclusions

This article analyzes online monitoring cases of PDs in three large power transformers using both the AE and UHF methods simultaneously. The transformers were selected for testing by the power system operator due to exceeding the typical concentration of hydrogen and acetylene dissolved in oil, indicating the potential presence of high-energy PDs. On the basis of the measurement results and statistical analyses obtained, the following conclusions were formulated:

• Simultaneous use of AE and UHF methods increases the effectiveness and reliability of PD detection. These methods are based on different physical phenomena (AE method records elastic waves; UHF method records electromagnetic pulses) and are susceptible to varying sources of interference. A high percentage of cases in which both methods simultaneously recorded pulses allowed us to confirm with high probability the presence of real PDs and not acoustic or radio interference. Statistical analyses showed that the agreement of the results for ‘Transformer A’ and ‘Transformer B’ was 65.6% and 72.8%, respectively, while for ‘Transformer C’, it was lower (43%).
• Monitoring PDs using the UHF method with the function of recording PRPD patterns allows for identifying the type of insulation system defect. For all the tested transformers, characteristic PRPD patterns were recorded, indicating the occurrence of PDs.
• Using a multi-channel AE system allows for the localization of defects in individual transformer phases. This allowed one to determine the potential location of discharge sources already at the monitoring stage. For ‘Transformer A’, AE pulses were recorded near the on-load tap changer; for ‘Transformer B’, in phase L3 on the high-voltage side; and for ‘Transformer C’, in phase L2 on the low-voltage side. The data obtained significantly accelerated the procedure for the precise location of the discharge sources using the TDoA technique.

The discrepancy in results between the AE and UHF methods may be due to several factors: (i) the presence of interference specific to the given method (e.g., AE may record signals generated by rainfall, which do not affect UHF); (ii) incorrectly selected pulse detection threshold (too low of a threshold causes recording of interference; too high of a threshold prevents the detection of low-energy discharges); (iii) the presence of so-called ‘silent discharges’, which generate negligible elastic waves and may not be detected by the AE method (an example of a silent PD is the formation of carbonization trees on the surface of the pressboard [45]); and (iv) the specific location of the PD source or its large distance from the PD detectors (antenna UHF or AE sensor). Metal structural elements of the transformer (windings, cores, and tank walls) may block the UHF signal, and solid insulation (pressboard barriers, spacers, and strips) may dampen elastic waves generated by PDs.

In conclusion, the simultaneous application of the AE and UHF methods in online PD monitoring improves the accuracy of discharge detection and verifies its actual occurrence, reducing the risk of false alarms.