Power electronic devices are widely used in renewable power systems, and the characteristics of partial discharge (PD) signals under repetitive impulse excitation are vastly different from those observed under power frequency excitation, causing many pattern recognition methods to fail. This paper proposes a improved identification method based on representation learning that can directly process PD sequences received by ultra-high frequency (UHF) sensors. First, a representation learning architecture based on the Transformer Encoder module is proposed and the model is pretrained in an unsupervised manner to extract features from the raw input discharge sequences. Then, the extracted features are processed using the ridge regression classifier to achieve end-to-end PD source identification. Finally, the effectiveness of the proposed model is validated through an experimental study, achieving an improved accuracy of 98.6%, which is better than classical PD identification deep learning models under repetitive impulse voltage. Furthermore, the model exhibits strong robustness against noise interference and sampling rates. The effectiveness of the model was also tested on an inverter-fed motor stator device. To further validate its practical applicability, it is essential to collect diverse PD data from various electrical equipment.

High-frequency repetitive pulse excitation is prone to induce partial discharge (PD) of power equipment, resulting in premature insulation failure. PD detection plays an important role in insulation condition monitoring. However, the severe switching electromagnetic interference (EMI) pulses of power electronic devices and the concentrated PD behavior pose difficulties for accurate PD detection. A novel method for detecting weak signals of PD pulses with EMI based on time-reassigned multi-synchrosqueezing transform (TMSST) and similarity distance measurement is proposed in this paper. Firstly, The TMSST method is employed to reassign the signal energy distribution in the time-frequency domain, which can significantly enhance the time-frequency resolution of pulse signals. Then, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering method based on the time-frequency energy distribution is employed to detect individual pulse signals. A mask-filtering technique is utilized to filter the noise and reconstruct the pulse signals. Next, the similarity of each extracted pulse signal is measured using the complexity-invariant distance (CID) criterion to identify and remove the EMI pulses. Finally, the performance of the proposed method is verified through simulated data and experimental tests under repetitive pulse excitation. The results demonstrate that the proposed method in this paper exhibits good performance in handling weak PD pulse signals and effectively eliminating the switching EMI pulse.

Partial discharge accelerates insulation degradation and is a major contributing factor to insulation failures in electrical equipment. Effectively detecting partial discharge signals and accurately identifying discharge types are crucial for assessing insulation conditions. However, in multi-source partial discharge scenarios, signal overlap in the time domain significantly complicates diagnosis. Furthermore, the characteristics of UHF signals received by sensors are influenced by both the type of insulation defects and the propagation path between the discharge sources and the sensors. Therefore, incorporating prior localization information is essential for accurate discharge type identification. This paper addresses the challenges of multi-source partial discharge type identification under repetitive pulse excitation by proposing a deep learning model. First, the Time-Reassigned Multi-Synchrosqueezing Transform (TMSST) is employed to mitigate the impact of signal overlap on feature extraction. Robust features are then extracted using the ResNeSt model with its Split-Attention mechanism. Considering the influence of UHF propagation paths on the received signal characteristics, prior localization information is encoded as auxiliary features and integrated with the time-frequency features. Finally, the Transformer Encoder backbone is utilized to effectively capture dependencies between time-frequency features and prior localization knowledge, enhancing both accuracy and interpretability. Experimental results demonstrate that the proposed model achieves an mAP of 98.7%, outperforming classical deep learning methods, and exhibits strong robustness to noise, highlighting its potential for engineering applications in multi-source partial discharge identification.

The directional multiple signal classification (Dir-MUSIC) algorithm based on the antenna gain array manifold has been proposed to find the direction of the partial discharge (PD) source in substations. However, PD signals are wideband signals and the antenna gain pattern functions are always different at different frequencies; therefore, the accuracy can be improved using a wideband Dir-MUSIC algorithm. In this paper, wideband Dir-MUSIC algorithms are discussed and a novel wideband Dir-MUSIC algorithm using the strength proportion (DirSP) is proposed. This algorithm estimates a focusing PD signal at a certain frequency using the strength proportion among different directions, and then the Dir-MUSIC algorithm can process the focusing PD signal at this frequency. In simulations, when the antenna gain functions among different frequency bins are quite different, the Dir-MUSIC algorithm loses accuracy; meanwhile, DirDP performs very well. In the experiments, we deal with six sets of samples, and the mean error and standard deviation are both smaller than $4°$ better than other methods.

Inspection robots are widely used in the field of smart grid monitoring in substations, and partial discharge (PD) is an important sign of the insulation state of equipment. PD direction of arrival (DOA) algorithms using conventional beam forming and time difference of arrival (TDOA) require large-scale antenna arrays and high computational complexity, making them difficult to implement on inspection robots. To address this problem, a novel directional multiple signal classification (Dir-MUSIC) algorithm for PD direction finding based on signal strength is proposed, and a miniaturized directional spiral antenna circular array is designed in this paper. First, the Dir-MUSIC algorithm is derived based on the array manifold characteristics. This method uses strength intensity information rather than the TDOA information, which could reduce the computational difficulty and the requirement of array size. Second, the effects of signal-to-noise ratio (SNR) and array manifold error on the performance of the algorithm are discussed through simulations in detail. Then, according to the positioning requirements, the antenna array and its arrangement are developed and optimized. Simulation results suggested that the algorithm has reliable direction-finding performance in the form of six elements. Finally, the effectiveness of the algorithm is tested by using the designed spiral circular array in real scenarios. The experimental results show that the PD direction-finding error is $3.{39}^{°}$, which meets the need for partial discharge DOA estimation using inspection robots in substations.

Extended exposure of power equipment to high voltages can lead to insulation failures. Accurate localization of partial discharge (PD) sources is essential for identifying these vulnerabilities effectively. Most deep learning methods are constrained to single-source scenarios or lack high localization resolution for multiple PD sources, leading to limited practical accuracy. To address these challenges, this paper proposes a novel deep learning-based method for Multi-DoA estimation of PD signals. Leveraging the theoretical foundation of the directional multiple signal classification (Dir-MUSIC) algorithm, the Temporal and Spatial Decomposition-Mixing Mechanism Network (TSDMNet) is introduced. This network captures temporal and spatial dependencies, enabling end-to-end processing of array-received data to identify the Multi-DoA of PD sources. Experimental results show that the proposed method outperforms classical deep learning models, particularly in low signal-to-noise ratio environments. For instance, at 0 dB, the root mean square error of direction estimation is less than 1 degree, demonstrating the model’s robustness and accuracy. Furthermore, the proposed TSDMNet eliminates the need for prior knowledge of source numbers or array geometry, enhancing its practical applicability.