Robotics and Deep Learning Framework for Structural Health Monitoring of Utility Pipes

A critical modern-day challenge for utility operators is condition monitoring of underground sewer infrastructure. Existing industry standard for underground sewer line inspection is based on sending a wire-guided robot with a closed circuit television (CCTV) camera through a pipe. A trained operator observes the video feed from the camera, and annotates it to record defects such as cracks, sags, offsets, root infiltrations, grease build up, and lateral protrusions. The success of a CCTV based robot system depends on visual observation and alertness of the operator. There is a likelihood that the operator fatigue and distraction may lead to missed observations. The CCTV based systems are expensive and man-hour intensive. We propose a deep learning based method to make the defect detection process automated without the need for an onsite operator to visually observe the video. The system is based on passing an autonomous camera-mounted robot through the pipe. The recorded video is analyzed using deep learning based algorithms. Our initial focus is to detect presence of cracks in polyvinyl chloride pipes, which are industry standard for sewer installations. We propose a deep learning framework including network architecture to detect presence or absence of a crack in a pipe sample. We also collect empirical data using an autonomous robot during laboratory trials to validate our approach. The data analysis indicates an accuracy of 89.42% in training and 83.3 % in validation. Further data collection and analysis is currently in progress and results will be reported in future.