Recurrence Plot Aided Partial Discharge Detection Framework Employing HFCT Sensor and Customized Convolutional Neural Network

In this article, a deep learning framework for automated detection of partial discharge (PD) events employing signature of a high-frequency current transformer (HFCT) sensor is proposed. For this contribution, one cycle PD signals captured using HFCT sensors were initially pre-processed and converted to RGB images using recurrence plot (RP), which can capture nonlinearity and dynamic fluctuations present in PD signals. The PD signal converted images using RP were then fed to a proposed customized lightweight CNN model for classification of different standard PD-types. The efficiency of the proposed method was initially validated on five emulated PD sources in the laboratory. In addition, experiments were also carried out on real-life insulators. Investigations revealed that the proposed RP aided customized CNN-based PD defect recognition method has delivered reasonably accurate results in classifying both the laboratory emulated as well as real-life PD sources. Compared to conventional pre-trained CNN models, namely AlexNet, GoogleNet, ResNet50 and MobileNet, the proposed CNN model is computationally inexpensive, with lesser number of learnable parameters and delivered accurate results at a significantly reduced training time. The proposed method can be implemented for automated detection of PD sources.