Deep Learning Based Navigation System for Train Line Equipment Inspection Using Drones
Yuki Hosono, Nico Surantha · 2024
Early damage detection and regular train overhead line equipment inspection are essential for safe and reliable train operation. Traditionally, these inspections have been conducted directly by power line engineers at night. However, such methods face challenges like a shrinking workforce and limited mobility. Attempts have been made to enhance mobility by installing cameras on test vehicles, yet the analysis of captured images still requires manual intervention. With the advancement of drones, high-resolution cameras, and deep learning technology, AI-based inspections using drones have become a promising solution. This study aims to develop an AI-based drone navigation system for automated inspection of train overhead line equipment, focusing on improving inspection efficiency and accuracy through model training for future drone deployment. In this study, a simulator was employed to generate images of overhead line facilities, and a machine learning model was trained using these images, based on the concepts of Dronet and FlowDronet from previous studies. The model's performance was evaluated using varying numbers of training images. A test dataset was created from images captured by a drone in flight and processed with optical flow techniques. The results revealed that the model trained with 22000 images achieved the highest performance, with an Average accuracy of 87.90%, a Collision Recall of 89.60%, and a Non-Collision Recall of 85.99%. These findings underscore the importance of larger datasets in enhancing model generalization capabilities and demonstrate the effectiveness of utilizing drones and AI for automated inspection of train overhead line equipment.