ABSTRACT
When a vehicle crosses a road bridge, the bridge bends and deforms in response. The dynamic characteristics are unique to each bridge and have been utilized for nondestructive health monitoring. Recently, we have developed a comprehensive system for remote monitoring, which is a platform of the Internet of Things era for the inexpensive and efficient monitoring of road infrastructures. In addition, we have proposed modeling techniques for existing bridges based on deep convolutional neural networks (CNNs). Our solution requires less heuristics and no optimization by hand; instead, it requires a camera and several sensors installed on the bridge components. By analyzing traffic situations using surveillance video data, we have collected a large amount of training data about passing vehicles, including traveling speeds, loci, and axle positions. The training data have been utilized for training a deep CNN. The CNN has brought rich information about passing vehicles and bridge dynamics, although the CNN accepts strain data collected by only a single strain gauge. We named this solution deep sensing, and its applications include vehicle weighing (bridge weigh-in-motion) and nondestructive bridge health monitoring based on anomalous signal detection. In actual health monitoring, we need to handle various types of heterogeneity due to constraints such as the deployable locations of strain gauges, as well as the types of bridges. In this paper, we present our experiences in dealing with the heterogeneity and discuss the practical system architecture, including interfaces, subsystems, and requirements for a fully automated modeling system for real road bridges. Our proposal contributes to the progress toward low-cost and compact health monitoring systems.
