Adversarial Spiral Learning Approach to Strain Analysis for Bridge Damage Detection.

When a vehicle passes over a bridge, the bridge distorts in response to the vehicle's load. The response characteristics may change over time if the bridge suffers damage. We consider the detection of such anomalous responses, using data from both traffic-surveillance cameras and strain sensors. The camera data are utilized to treat each vehicle's identified properties as explanatory variables in the response model. The video and strain data are transformed into a common feature space, to enable direct comparisons. This space is obtained via our proposed spiral learning method, which is based on a deep convolutional neural network. We treat the distance between the video and strain data in the space as the anomaly score. We also propose an adversarial unsupervised learning technique for removing the influence of the weather. In our experiments, we found anomalous strain responses from a real bridge, and were able to classify them into three major patterns. The results demonstrate the effectiveness of our approach to bridge damage analysis.