A Kriging Surrogate Model for Structural Health Monitoring of Miter Gates in Navigation Locks

Abstract

Structural health monitoring (SHM) plays a significant role in monitoring the performance of the inland waterways civil infrastructure such as navigation locks. The United States has s $500B replacement value of its inland waterway infrastructure. Unexpected shutdowns of miter gates in navigation locks caused by any kind of damage produces tremendous economic loss, particularly if the shutdown is unplanned. Sensing is the first step in any process of damage detection strategy that collects data from sensing network. Lock gates are normally instrumented with strain gauges. A missing contact (or gap) between the concrete supporting wall and gate, one of the most common failure modes of lock gates, creates undesirable overloads and produces different levels of damage. An inverse model is needed to map the data extracted from the strain gauges to damage-sensitive features that are used to build degradation models and make decisions. An inverse Bayesian analysis is performed to identify damage by using strain gauge data. Although the posterior distribution estimations could provide probabilistic solutions to the damage identification, these models usually have high computational cost, and thus they are not suitable of performing real-time health monitoring. In this preliminary work, a Kriging or Gaussian process regression (GPR) inverse model is designed as a surrogate model of the finite element model of the Greenup lock miter gate, to map the strain gauge data to the damage size (i.e. gap length in miter gate contact). Hydrostatic loading scenarios governed by upstream and downstream heights with varying gap length are considered in the design of the Kriging model. Both training and testing data sets are predicted using the designed model. Overfitting problem caused by high nonlinear behavior of the data are solved. The results illustrated that Kriging, as a surrogate model, is a fairly reliable replacement to the computationally expensive inverse finite element model in damage identification. In addition, prediction performances using other Kriging models with different basic covariance functions and sizes of the training data set are compared.