Deep Learning Approach for Noninvasive Tracking Detection in Silicone Rubber Insulators Using Fluorescent Fiber Sensing

This study presents a non-invasive approach for early detection of tracking phenomena in silicone rubber insulating materials during tracking studies, adopting IEC 60587 standards, using an optical fluorescent fiber sensor and a deep learning technique. The sensor captures light emissions during surface discharge, providing a high-resolution, noise-free signal without electrical contact. The signal is transformed into Gramian Angular Field (GAF) images representing temporal correlations to classify the discharge stages effectively. A novel restructuring strategy is employed to extract and reorganize diagonal sub-images from the GAF matrices, reducing the redundancy while emphasizing the key temporal features. These reorganized GAF inputs train a convolutional neural network (CNN) model to classify four distinct discharge stages. The model is evaluated with both single-channel and dual-channel GAF representations. Experimental results show that the proposed method achieves 98.22% classification accuracy, outperforming raw GAF inputs and demonstrating strong potential for real-time condition monitoring of high voltage insulation systems.