Damage detection in structural systems utilizing artificial neural networks and proper orthogonal decomposition

• Published in: Structural control and health monitoring Vol. 26; no. 2; pp. e2288 - n/a
• Main Authors: Eftekhar Azam, Saeed, Rageh, Ahmed, Linzell, Daniel
• Format: Journal Article
• Language: English
• Published: Pavia Wiley Subscription Services, Inc 01-02-2019
• Subjects: artificial neural network; Artificial neural networks; Bridge loads; classification; Computer simulation; Damage detection; Decomposition; Finite element method; Learning theory; Neural networks; Proper Orthogonal Decomposition; Railroads; Railway bridges; Railway stations; regression; strain; Structural damage; Three dimensional models; Trains; Truss bridges
• ISSN: 1545-2255; 1545-2263
• DOI: 10.1002/stc.2288

Summary
A supervised learning scheme is proposed for detecting, locating, and quantifying the intensity of damage in structures using Artificial Neural Networks (ANNs) and Proper Orthogonal Decomposition (POD). For structural systems, such as buildings and bridges, Proper Orthogonal Modes (POMs) associated with their response are functions of (1) applied external loads and (2) mechanistic properties. In the present research, a supervised learning strategy was adopted to help discriminate POM variations because of damage from damage caused by applied load variations. A neural classifier was trained to categorize response to different load patterns, and a regression ANN was subsequently trained using an ensemble of applied loads to detect possible damage from the categorized POMs. To demonstrate the effectiveness of the proposed approach, simulated experiments were performed with the intent of identifying damage indices for a railway truss bridge. A validated, three‐dimensional (3D) finite element (FE) model of an existing bridge was used to generate strain time histories under train loads measured from weigh‐in‐motion (WIM) stations near the bridge. The efficacy of the proposed method was demonstrated through these simulated experiments.