Generative adversarial network for predicting visible deterioration and NDE condition maps in highway bridge decks

Abstract

Bridge decks tend to degrade faster than other bridge components due to environment exposure and vehicular loading. Periodic degradation monitoring is needed for timely rehabilitation measures and development of service life models in bridge decks. Surface degradation are identified through visual inspection (VI) and post-processing of high-definition imagery. Although VI is the primary NDE method employed by most transportation authorities, many anomalies (e.g., cracking, corrosion, and delamination) remain hidden under the surface until deteriorations have grown large enough to surface (e.g., spalling). Subsurface degradation is best identified through other forms of non-destructive evaluation (NDE). Inferences can be made between the various NDE methods, as the mechanisms behind the damages sensed by each method are shared. For instance, condition map from an NDE method may infer future visible deterioration, as well as condition maps for other NDE methods. In this paper, a deep learning approach based in a conditional generative adversarial network is presented for modeling of plausible visible deterioration and NDE condition maps. Two applications are explored: (i) visualization of plausible future deterioration based on current NDE condition map, and (ii) visualization of condition maps for NDE methods from other NDE methods. Field and experimental data from the BEAST facility at Rutgers University are used to develop the training databases for each application.