Deep learning- and infrared thermography-based subsurface damage detection in a steel bridge

The aging and deterioration of bridge infrastructure is becoming a serious issue around the world. In this research, a new deep-learning-based method is proposed to detect subsurface damage in the steel elements of a bridge using thermography, without physical contact. Thermal images of the structural steel elements on the Arlington Bridge in Winnipeg, Manitoba, were captured using an uncooled microbolometer and were then used to train and validate a deep inception neural network; a maximum testing accuracy of 96% was achieved. Next, ultrasonic pulse velocity tests were conducted for validating thermal-image-based subsurface damage detection, and a contour map of the sections was plotted and compared with a deep-learning-based results. This research demonstrates that the combination of infrared thermal technology with a deep neural network is a practical approach to autonomously detecting subsurface damage in the elements of a steel bridge, with minimum human intervention.