TOWARD POPULATION-BASED STRUCTURAL HEALTH MONITORING FOR SHEAR-CRITICAL PRESTRESSED CONCRETE GIRDERS

Abstract

The increased interest in resilient bridge infrastructure has raised interest in improved methods for analyzing structural integrity. The need for recommendationson bridge load postings, retrofitting, or replacements motivates an active application of Structural Health Monitoring (SHM). SHM has been developed over the past few decades to assess the condition of structures, such as bridges. One of the main challenges for achieving an accurate assessment is the lack of failure data. This is the case for shear failures in prestressed concrete girders. Population-based SHM (PBSHM) is a subfield of SHM and Machine Learning (ML) that overcomes the lack of failure data by adapting information from other structures within a population to the structure of interest. PBSHM was invented at the University of Sheffield, U.K., by the Dynamics Research Group (DRG). This study presents a pioneering application of PBSHM to shear failure load prediction of pretensioned prestressed concrete girders by directly working with DRG in a proof-of-concept study. To train an ML model to predict shear failure load more accurately than any single shear design procedure alone, the team collected data on the well-established shear design procedure predictions, which are the inputs, and on the experimental girder shear failure loads, which are the outputs. This new input-output pairing is our intended contribution because a typical prediction (regression) problem in ML would use the involved material and geometric properties as the input instead, bypassing any of the existing design procedures. Since design procedures contain well-established engineering knowledge, this ML application incorporates interpretability and informativeness (I2 ), an overarching guiding principle from a computer science expert on the team. The multidisciplinary work of this study involves both structural engineering and ML knowledge; we are fortunate to have a structural engineering expert on the team. We either obtained or calculated predictions from suitable design procedures and matched their corresponding experimental shear failure loads for problem formulation and ML training and validation. We first encountered a few unexpected and nontrivial challenges, however we eventually emerged with strategies to deliver this proof-of-concept study. We came up with a small collection of carefully prepared datasets, a simple yet justified problem formulation, and an application of an interpretable machine learning (IML) technique to design and initialize multilayer feedforward neural network (FFNN) for the regression problem. Preliminary training and validation results are reported and call for further improvement. Future work has been identified for continued and more productive PBSHM effort.