Deep Bayesian neural networks for damage quantification in miter gates of navigation locks

Abstract

Inland navigation infrastructure like locks and dams form a vital part of the global economy. Locks facilitate the transport of hundreds of millions of dollars’ worth of goods on a daily basis. A primary cause for downtime of locks in the United States is damage to lock gates. Current inspection methods involve the complete closure of locks to visually inspect for damage. A common target of such inspections is the identification of “gaps” that form along the bearing surface boundary of miter gates. These gaps accelerate the fatigue failure of the gate by disrupting the designed load distribution mechanism. This article presents a novel engineering application of structural health monitoring for full-scale civil infrastructure with a method to automatically quantify the damage quantity of interest, that is, the gaps using measured strain data. We propose a framework for damage estimation of full-scale civil infrastructure in general and miter gates in particular, leveraging recent advances in deep Bayesian learning. A new two-term loss function is produced to increase the accuracy of the trained networks and the model uncertainties are conveyed using Monte Carlo dropout. In addition, we propose a strategy to model bearing surface gaps using non-linear contact analyses and use the proposed model to determine the sensitivity of measured strains to damage. The proposed framework is implemented for the miter gates at the Greenup locks and dam. Finally, the proposed methodology is validated using measured data. Slopes measured from the lock gate are used as the input to the trained networks to estimate the gap depths. The finite element model is updated using the estimated gap depths. The predicted slopes and strains from the updated model are shown to match the measured strains and slopes well. The results demonstrate the efficacy of the approach for damage detection in full-scale civil infrastructure.

Model correction and updating of a stochastic degradation model for failure prognostics of miter gates

“…The formation of this gap is due to the contact degradation between the quoin block attached to the gate and the quoin block attached to the wall that supports the gate laterally. The formation of the bearing gap can be detected using sensor data or from features derived from this data [2,19,[30][31][32]. The term is the derived transition matrix obtained from the stochastic degradation model.…”

A Novel Framework for Integration of Abstracted Inspection Data and Structural Health Monitoring for Damage Prognosis of Miter Gates

“…The formation of this gap is due to the contact degradation between the quoin block attached to the gate and the quoin block attached to the wall that supports the gate laterally. The formation of the bearing gap can be detected using sensor data or from features derived from this data [2,19,[30][31][32]. Figure 6 idealizes the loss of contact in the physical-based FE model and shows the top view of the quoin blocks.…”

Efficient development of vision-based dense three-dimensional displacement measurement algorithms using physics-based graphics models

“…29 Steps to create a PBGM for vision-based displacement and/or strain measurement are illustrated in Figure 2. First, an FE model of the target structure is created (FE models are sometimes available for civil structures with high significance, for example, miter gates maintained by the US Army Corps of Engineers 27,28 ). Then, cross sections and surface textures are added using a graphics modeling software to convert the model to the photo-realistic one, that is, PBGM of the undeformed structure.…”

“…18,[24][25][26] Despite the success of these approaches for tracking multiple points, the extension of the methods to the estimation of structural properties of more complex structures at higher spatial resolution is not straightforward (e.g. detailed FE model of a miter gate can have millions of degrees of freedom, 27,28 and manually defining tracked points and two-dimensional (2D) to 3D conversions for all visible FE nodes/elements is impractical).…”