MS-FNO for Damage Localization and Quantification on a Steel Truss Bridge

Abstract

Neural operators have emerged as powerful tools in structural health monitoring for their ability to map between function spaces. This study introduces the ModalStiff Fourier Neural Operator (MS-FNO), designed to tackle inverse problems like localizing and quantifying structural damages. MS-FNO learns mapping from modal shape spaces (vertical modes of deck) to stiffness spaces (stiffness of tension member). The validity of MS-FNO is validated on both finite element dataset (FE-Set) and real-world experiment dataset (EXP-Set) in which the FE-Set is created by simulating random variations in the stiffness of tension members within a 3D FE model. First, MS-FNO is trained on FE-Set and tested on scenarios of full cutting each tension member. Subsequently, MS-FNO is fine-tuned with intact data from EXP-Set and evaluated on a damaged scenario involving a full cut at the 5/8th span. Compared with existing widely used modal shape-based method like COMAC, MS-FNO demonstrated superior accuracy in localizing and quantifying damages, despite challenges such as sensitivity variances in modal shape and misidentification among adjacent members. Post-finetuning, although significant errors were observed in undamaged members, MS-FNO effectively localized and quantified damage in affected areas.