An intelligent framework of upgraded CapsNets with massive transmissibility data for identifying damage in bridges

Abstract

Structural monitoring systems installed on bridges are capable of capturing large-scale dynamic responses online and in real-time. The response data of the bridge under different loading conditions is used for condition assessment of the bridge. There is a chasm between monitoring data and damage assessment due to the difficulty in revealing the relationship between the massive monitoring data and damage. this study proposes an intelligent framework for identifying (localizing and quantifying) damage in a bridge from massive acceleration responses. The framework features (i) physical big-data quantity and (ii) an intelligent identification model. The former specifies a mechanical quantity termed damage pattern spectrum of power-spectrum-density transmissibility (PSDT), the PSDT-damage pattern spectrum for short, constructed to protrude damage information from massive acceleration responses; the latter specifies an upgraded Capsules Networks (CapsNets), Up-CapsNets for short, which intelligently locate and quantify damage via deep learning with inputted PSDT-damage pattern spectrum. The proposed intelligent framework is verified through a numerically truss bridge model and an experimental suspension bridge model. The results demonstrate that this intelligent framework can identify locations and severity of damage with greater accuracy, stronger immunity to noise, and higher generalization than existing convolutional neural networks (CNNs) and CapsNets. Such a peculiarity can be attributed to two points: (i) the sophisticated function of the Up-CapsNets in recognizing damage patterns over conventional deep learning networks; and (ii) the unique qualification of PSDT-damage pattern spectrum in refining damage information over raw acceleration responses. The proposed intelligent framework provides a viable approach for monitoring the condition of a bridge by combining deep-learning deduction and big-data dynamic responses.