Simultaneous classification and localization of partial discharges in gas insulated switchgear using recurrent neural network

Introduction

Classification and localization of the sources for partial discharge (PD) activities are essential for structural health monitoring and maintenance of gas insulated switchgear (GIS) equipment. Even though several quality assurance checks are performed in factory, a variety of defects might appear during production, shipment, and installation of GIS components and these defects can initiate PDs. Even though the magnitude of PD pulses is small, they induce a cumulative degradation, which eventually result in GIS failure. PDs are associated with several attributes such as vibration, light emission, decomposition of sulphur hexafluoride (SF₆) gas and electromagnetic (EM) radiation, etc. PD diagnosis is possible by identifying these detectable quantities. The IEC 60270 and IEC 62478 standards,^1,2 respectively, define the conventional methods (such as current and voltage measurement sensing) and nonconventional methods (such as optical, chemical, acoustical, and EM measurement sensing) for PD activity monitoring in electrical equipment. The conventional PD monitoring methods often require galvanic connections and are susceptible to external electrical noise, limiting their effectiveness in noisy high voltage (HV) environment like GIS. These methods also face challenges in installation, especially in sealed systems, and are not ideal for continuous online monitoring. Among the nonconventional methods, ultra high frequency (UHF) technique is widely used for capturing EM emissions due to PD, as EM wave is insensitive to temperature and pressure variations, propagates through all dielectric medium and are suited for continuous real time monitoring.^3,4 PD events in GIS release EM energy in UHF regime of 0.5–3 GHz and antennas designed to capture the EM radiation in the UHF band can be used as sensors for PD detection. Each type of PD signal exhibits its own signature and attributes to varying degrees of damage to GIS insulation. Therefore, PD signal detection, classification, and localization are important for assessing the health of the dielectric insulation and developing timely maintenance strategies. A review of the literature shows different network architectures for classification or localization but there is little work on simultaneous classification and localization of PD sources. In, ⁵ a collaborative domain adaptation (CDA) network method is proposed for localization of PD sources utilizing maximum mean discrepancy (MMD) and adversarial training for two-domain alignment after noise filtering using synchronous squeezing wavelet transform (SWT) and feature extraction using 1-dimensional attention based convolutional neural network (1DACNN). A domain-invariant long short-term memory (LSTM) is proposed for localization in, ⁶ following noise filtering using wavelet packet transform, with a CDA module employing MMD-based domain adaptation to enable knowledge transfer from source to target domain. In, ⁷ a dual-task network (DTN) is proposed using an attention based bidirectional gated recurrent unit (ABiGRU) for extracting temporal dependencies from PD signals, while a multi-gate mixture of experts (MMoE) layer facilitates learning between the diagnosis and localization modules. A Multi-Task Learning Network (MTLN) for evaluating PD conditions in GIS is proposed in,^8,9 enabling simultaneous diagnosis, localization, and severity assessment. By combining transformer and LSTM models, the approach captures temporal and spatial dependencies in PD signals. A novel Multi-pathed Hierarchical Mixture-of-Experts (MHMoE) mechanism allows expert units to specialize in individual tasks while benefiting from shared information. The model achieved 98.36% diagnosis accuracy, 94.63% severity assessment accuracy, and a localization error of 7.96 cm. A comprehensive survey of the most relevant research on the detection, classification, and localization of partial discharge (PD) sources using the radiometric or ultra-high frequency (UHF) method is detailed in. ¹⁰ A common limitation in the approaches listed in,^5,6 is the need for prior signal denoising using wavelets. Moreover, the network architectures listed in^5-9 employ domain adversarial training (DAT) to transfer knowledge from simulation samples (source domain) to real-world PD localization (target domain).

In this work, we present a simplified network architecture using recursive neural network (RNN) without the need for prior noise filtering which can simultaneously classify and localize PD signals directly in the target domain. The organization of this work is as follows: UHF sensing inside GIS is introduced, followed by data acquisition, labeling for different defect types and location, and RNN training and testing of the network performance for classification and localization of PD defects inside the L-GIS section.

UHF sensing in GIS

A typical GIS is made up of straight, L and T sections filled with SF₆ gas, and separated by solid epoxy insulating discs. Figure 1 shows the simplified cross section of single-phase straight GIS section. It consists of an inner conductor surrounded by coaxial outer conductor (enclosure), insulating spacers, handholes, and internal PD monitoring sensor. The inner conductor carries high voltage (HV), while the outer enclosure is grounded, and the space between them is filled with SF₆ gas. The inner conductor is supported by insulating spacers made of casting epoxy. The insulating spacers can also be employed as dielectric windows for PD monitoring. Handholes are provided to place sensors internally for monitoring PD events. UHF antennas reported for PD detection typically have large bandwidth^11-13 for capturing the short duration EM radiation occurring during PD. Due to multipath propagation and higher order mode generation inside GIS, ¹⁴ UHF antenna with circular polarization is used in this study. ¹¹

OPEN IN VIEWER

Figure 1.

A mid-section view of straight GIS section illustrating different PD defects, and implementation of internal and external UHF sensors for PD sensing.

Experimental setup and PD measurements

Figure 2(a) shows the experimental setup for simulating different types of PDs inside GIS. The setup consists of a HV transformer (50 Hz, 0–100 kV, 10 kVA), capacitance divider, test cell with defined electrode configurations for simulating different types of PDs, a 3 m long L-GIS section, two identical cosine slot Archimedean spiral antenna (CSASA), ¹¹ and a digital storage oscilloscope (DSO) (Teledyne LeCroy Wave Runner 604Zi). The CSASA sensors, one before the L-bend (UHF1) and one after the L-bend (UHF2) were placed in the respective GIS hand holes at 50 cm distance from the shorting plates and were connected to the DSO using two identical low loss coaxial cables. Commonly occurring defects, corona on HV (C), particle movement (P) and surface discharge (S) were considered. Corona on HV was generated using a needle plane electrode configuration as shown in Figure 2(b). The test cell consists of a HV needle electrode of 50 μm radius of curvature and a circular bottom ground plane electrode of 55 mm diameter. The gap between the needle tip and ground electrode was maintained at 5 mm. Particle movement discharges were generated using metallic contaminants shown in Figure 2(c). The test cell consists of a 25 mm diameter spherical upper HV electrode and a 55 mm diameter concave shaped bottom ground electrode separated by 10 mm with a 1 mm diameter Aluminium spherical ball placed on the ground electrode to simulate PDs due to PM. Surface discharges were initiated at the triple junction of the HV conductor, epoxy spacer, and SF₆ gas using an IEC(b)-type HV electrode and grounded plane configuration shown in Figure 2(d). A 1.5 mm thick epoxy insulating disc was sandwiched between the electrodes and HV is supplied to IEC(b) electrode to initiate SD type of discharges.

OPEN IN VIEWER

Figure 2.

(a) Experimental setup for classification and localization of PD signals inside GIS using CSASA sensors; electrode configurations in SF₆ filled test cell for (b) corona on HV, (c) particle movement and (d) surface discharge types of dielectric insulation defects.

The defects were simulated in test cell configurations illustrated in Figure 2 with the high-voltage (HV) electrode and grounded bottom electrode separated by a 5 mm thick acrylic insulating tube of 50 mm height. The enclosure was filled with SF₆ gas pressurized to 3 bar to emulate operational GIS conditions. PD induced UHF signals initiated at the inception voltage for each PD test cell positioned inside the L-GIS section were collected by the CSASA sensors and stored in DSO at a sampling rate of 10 GHz. Table 1 presents the inception voltages required to generate PDs due to simulated defects in test cells with defined electrode configurations discussed above. The inception voltages tabulated for each PD defect in Table 1 is an average of 10 measurements at 3 bar SF₆ pressure. For each of the three defect types, PD experiment was conducted at 8 different positions inside the GIS by placing the corresponding test cell at distances of 15, 30, 45, and 60 cm from both ends of the shorting plates and PD signals were acquired by the two CSASA sensors (UHF1 and UHF2).

OPEN IN VIEWER

Table 1.

Mean inception voltages of defects inside test cell filled with SF₆ gas at 3 bar.

PD Type	Inception Voltage (kV)
Corona on HV	11.5 ± 0.6
Particle Movement	8.6 ± 0.8
Surface Discharge	11.2 ± 0.6

As PD signals are non-stationary in nature, 1100 signals were collected for each PD defect type and location by UHF sensors 1 and 2 separately. Thus, 26,400 (1100 × 8 × 3) signals were gathered by UHF1 and UHF2 individually using the GIS measurement setup, totaling to 52,800 signals for network training and assessment. Each set of measurements recorded by UHF sensors 1 and 2 was labeled using a naming convention in the format of “XYZ”, where, X denotes the type of discharge which can be either Corona (C), Particle Movement (P) or Surface Discharge (S), Y denotes the distance from the shorting plate which can be either 15, 30, 45 or 60 cm, and Z denotes the shorting plate from which the distance is measured. If the distance is measured from shorting plate 2 an apostrophe is added else, it is left empty. In total, 24 unique labels were created for the PD induced UHF signal dataset gathered for the 3 PD defects located at 8 different positions inside the L-GIS. PD induced UHF signals acquired from the L-GIS section are presented in Figure 3. Figure 3(a) shows the PD signals recorded by UHF sensors 1 and 2 for test case “C15” corresponding to corona type defect on HV located 15 cm from shorting plate 1. Similarly, Figure 3(b) and Figure 3(c) depict UHF signals for test cases “P30” and “S45” representing particle movement and surface discharge at 30 and 45 cm from shorting plate 1, respectively.

OPEN IN VIEWER

Figure 3.

Pd induced UHF signals for different defect types gathered in L-GIS section. (a) Corona on HV conductor at 15 cm distance from shorting plate 1, (b) particle movement at 30 cm distance from shorting plate 1 and (c) surface discharge defect at 45 cm distance from shorting plate 1.

Recurrent neural network - LSTM

A RNN is a network architecture that can be used to solve classification or regression problems for time-series data. In practice, simple RNNs trained through backpropagation experience a problem with learning longer term dependencies due to either a “vanishing” or “exploding” gradient problem. ¹⁵ A special type of RNN that overcomes this issue is the LSTM network. However, standard LSTM processes input sequences in a single temporal direction and may overlook contextual dependencies that occur earlier or later in the sequence. To address this limitation, a bidirectional LSTM (BiLSTM), essentially two LSTMs joined together, one processing the input sequence in the forward direction and the other in reverse—is used in this study to improve network performance by capturing both past and future temporal dependencies more effectively. Simpler temporal models such as unidirectional LSTMs and GRUs were evaluated, however Bi-LSTM consistently outperformed them in classification accuracy and localization precision due to its ability to incorporate contextual information from both past and future time steps.

Figure 4(a) illustrates the RNN architecture employed for simultaneous classification and localization of PD-induced UHF signals. The network comprises a sequence input layer, followed by four BiLSTM layers in cascade, a fully connected layer, a softmax layer, and a classification output layer. The network in total consists of 8 layers with a total of 85,000 learnable parameters which are optimized through backpropagation using the training dataset. Hyperparameter tuning was performed empirically by evaluating combinations of hidden units, learning rates, and batch sizes through iterative experimentation to minimize validation loss and maximize classification and localization accuracy. The dataset was split into 80%, 10%, and 10% for training, validation, and testing respectively. To prevent overfitting, a validation set was used in conjunction with the training dataset every 50 iterations. An adam optimizer was used as a solver with the initial learn rate set to 0.01 and the learn rate schedule was set to “piecewise”. The learn drop period and drop factor was set to 4 and 0.5, respectively. The network was implemented using MATLAB's deep learning toolbox and trained on a workstation equipped with a 4.10 GHz Intel^® Xeon^® W-1250P CPU, 32 GB RAM, and a 6 GB NVIDIA RTX A2000 GPU. Training was performed with a mini-batch size of 256, 165 iterations per epoch, and a maximum of 20 epochs.

OPEN IN VIEWER

Figure 4.

BiLSTM based RNN for classification and localization of PDs in GIS. (a) Network architecture implementation, (b) training progress and (c) confusion matrix for testing.

Figure 4(b) shows the training progress, with the final training dataset accuracy at 98.44% and validation set accuracy at 92.20%. The training and validation losses are 0.0618 and 0.3556, respectively. Testing the trained network on the withheld 10% data yielded an accuracy of 95.95%. Figure 4(c) shows the confusion matrix for the true and predicted classes. The networks prediction was the highest for labels C45, P30, P45′ and the lowest for label S15′. The misclassification between S15′ and C60′ may be attributed to the similarity in signal features and propagation characteristics received at the UHF sensors as EM waves generated by PDs inside the metal-encapsulated GIS undergo multiple reflections from internal surfaces, resulting in the formation of several higher-order propagation modes and complex signal attenuation.

Discussion

The paper proposes a BiLSTM-based RNN for simultaneously classifying and localizing PDs within GIS. Unlike traditional time-based localization methods, this approach avoids the need for precisely extracting temporal characteristics from different PD signals. Time based approaches often encounter difficulties in non-convergence of the nonlinear equations associated with Time of Arrival and Time Difference of Arrival, especially in environment like GIS due to multipath propagation and higher order mode generation. ¹⁴ Adopting data driven techniques can significantly reduce the complexity of solving classification and localization in GIS but requires prior training with labeling. A comparison of the current network with CDAN proposed in ⁵ shows an improvement of around 5 cm in spatial resolution. The domain-invariant LSTM and DTN proposed in ⁶ and ⁷ have better spatial resolution at around 11 cm. However, the absence of separate noise filtering stage and localization directly in the target domain are advantages of this simplified network architecture. A future scope of the current work would be to treat localization as a regression problem instead of a classification problem to provide continuous measurements for identifying defect location.

Conclusion

Different types of PD signals that occur due to insulation defects inside GIS were classified and localized using a simplified customized RNN using BiLSTM layers. The model was trained with 42,240 signals corresponding to 3 defect types (1760 signals per defect per location). The model showed an overall accuracy of around 96%, indicating the feasibility of simultaneous PD defect classification and localization with a spatial resolution of 15 cm. To further improve localization accuracy, a regression-based approach is being explored to predict defect positions as continuous values, potentially enabling sub-15 cm spatial resolution.