Experimental verification of a vibration-based damage identification method in a timber structure using Neural Network ensembles
Title: Experimental verification of a vibration-based damage identification method in a timber structure using Neural Network ensembles
Author(s): U. Dackermann, J. Li, B. Samali, F. Choon Choi, K. Crews
Paper category : conference
Book title: RILEM Symposium on On Site Assessment of Concrete, Masonry and Timber Structures - SACoMaTiS 2008
Editor(s): L. Binda, M. di Prisco, R. Felicetti
Publisher: RILEM Publications SARL
Publication year: 2008
Pages: 1049 - 1058
Total Pages: 10
Nb references: 10
Abstract: Vibration-based damage identification methods utilise the abnormality in dynamic fingerprints of a structure to detect damage. Dynamic fingerprints can be extracted from time histories, frequency response functions, natural frequencies or modal strain energies. Damage occurring in a structure alters these dynamic fingerprints, and therefore they can be used as reliable tools to identify damage.
This paper presents an overview of a project that aims to identify structural damage by examining a variety of dynamic fingerprints. Artificial neural networks (ANNs) are developed to identify pattern changes associated with damage. Neural network ensemble techniques are adopted to fuse outcomes of individual network estimations and to provide a more accurate and reliable damage prediction.
In detail, a procedure is presented that utilises the damage index method, which is based on modal strain energy changes, to determine the location and the severity of single damage.
Numerical models of timber beams inflicted with several types of damage are generated. A laboratory timber beam damaged at mid-span is experimentally tested and analysed. Damage is identified by neural network ensembles that use damage index values as input patterns. The networks are first trained with indices of numerical timber beams and then tested with data obtained from the laboratory timber beam.
Online publication: 2009-05-26
Publication type : full_text
Public price (Euros): 0.00