BRIDGE SLAB ANOMALY DETECTOR USING U-NET GENERATOR WITH PATCH DISCRIMINATOR FOR ROBUST PROGNOSIS

Abstract

More than 50 years aging civil infrastructures have deteriorated, then structural diagnosis and periodic prognosis become critical for predictive maintenance. In terms of the bridge inspection every 5 years in Japan, we have collected a lot of human eye inspection. In context of digital structural monitoring, in addition to the past human inspection we make the most of drone flight images. However, human subjective judge includes individual bias, then a measurable objective score should be quantified using a unified anomaly distance from a health condition. Supervised learning, e.g. classification and semantic segmentation method is not always robust for unseen data. If we address the unlearned blind feature without any experience, prediction error might be a higher hurdle to overcome low precision and less recall problem. The generative anomaly detection via unsupervised learning has been growing in various fields, e.g. medical, manufacturing, food, and materials. If the distance and angle to the target damage interest could be controlled among a feasible range, and if the background noise could be removed and relaxed, then concrete surface damage and steel paint peel or corrosion would enable to discriminate them for predictive maintenance. In this paper, we propose a steel anomaly detector method to compute anomalous scores automatically, where we customize several U-shape skip-connected generator network with patch GAN discriminator. Exactly, we have create an encoder-decoder network using the VGG19 based U-Net generator with a patch discriminator. Furthermore, we explore robust unified anomaly score indicator for the target concrete and painted steel parts to analyze deterioration prognosis, so as to monitor the current status far from a health condition. Finally, focusing on the bridge slab, we exploit toward the inspection images with the number of 10,400, where they contains reinforcement concrete slab at 400 bridges under the direct control of national managers. In order to be stable learning and robust structural health monitoring, we demonstrate to visualize several anomalous feature map for precisely and full-covered digital inspection.