Nonintrusive Ultrasonic Sensing and Deep Learning for Outdoor Ceramic Insulator Assessment
Abdulla Lutfi, Ayman El‐Hag, Khaled Shaban
Abstract
Insulator failure, often triggered by contamination flashovers and punctures, poses significant risks, leading to substantial technical and economic losses. Ceramic insulators are particularly vulnerable due to their hydrophilic and brittle properties, which compound these risks, thereby impacting the reliability of the electrical grid. This study introduces a novel approach that utilizes ultrasonic sensors in conjunction with a Convolutional Neural Network (CNN) to detect and classify defects in ceramic insulators. The model is trained using ultrasonic signals obtained from a defective two-disc configuration under controlled laboratory conditions, then evaluated with a three-disc configuration across various settings, including a real-world 138/13.8kV substation environment. The model demonstrates high accuracy, achieving rates of 99% and 96% for laboratory settings, and 91% in field conditions. Furthermore, through the integration of Gradient-weighted Class Activation Mapping (Grad-CAM), the model’s decision-making process is explored, revealing a focus on higher frequency components within the ultrasonic spectrum. This insight underscores the potential of deep learning in enhancing non-intrusive insulator assessment techniques, paving the way for more reliable electrical grid operations.