Deep learning based structural health monitoring with autonomous UAVs

点击翻译按钮获取中文摘要

An automated reliable, efficient, and robust health monitoring system is urgently needed to detect damaged infrastructure in Canada's aging bridge system. Many vibration-based methods have been developed to detect bridge damage using contact sensors. However, these methods are unreliable and expensive, because they are vulnerable to uncertainties, noise, and environmental changes and require many installed sensors. Visual inspection by trained engineers is the current main approach, but is also costly, and biannual inspection cannot adequately prevent sudden collapse. Recently, computer vision (CV)-based damage detection methods have been proposed to support visual inspection. At present, this method is limited to detecting only one damage type. Additionally, it uses weak damage sensitive features extracted from traditional image processing and traditional machine learning to classify damage features. Thus, it is not robust to changes in the lighting conditions of input images. Since 2017, I have pioneered deep learning (DL)-based damage detection to overcome the limitations of traditional CV-based approaches through my previous NSERC Discovery Grant. This research has received worldwide interest, and numerous follow-up studies have been conducted. I also initiated the development of an autonomous drone flight method to enable flight in global positioning system (GPS) denied areas using an ultrasonic beacon system (UBS). These areas include the space beneath a bridge deck, where GPS is unavailable, and safety-critical areas, where important structural members and their connections are found. However, extensive investigation is required to develop a fully automated bridge inspection and management system. Such a system is the long-term goal of the proposed research program. The short-term objectives of this proposal are to (1) develop pixel-level multiple-damage identification based on advanced DL methods for external and internal damage using regular and thermal cameras, (2) develop an autonomous flight method for an entire bridge system without using GPS or UBS, (3) develop a holistic three-dimensional (3D) damage mapping method for efficient bridge management, and (4) integrate of all the developed methods for the fully automated bridge inspection system. This automated inspection system will be reliable, accurate, and cost-efficient by providing an explicit, visual, and holistic 3D damage map for efficient bridge management. Automation enables frequent inspection and can prevent sudden collapse by detecting early-stage internal and external damage. This research will produce a prototype of the automated bridge inspection system, which will be relevant to artificial intelligence and mechanical and electrical industries worldwide. The prototype will be developed using a DL method and the autonomous flight drone system with wireless auto-charging. Accordingly, University of Manitoba continues to be a world leader in this research topic.