Feasibility of damage detection of Tsing Ma Bridge using vibration measurements

In this paper, we study the feasibility of vibration-based damage identification methods for the instrumented Tsing Ma Suspension Bridge with a main span of 1377 m. Emphasis is placed on how to deal with the noise-corrupted/uncertain measurement data and how to use the series data from the on-line monitoring system for damage detection. Numerical simulation studies of using the noisy series measurement modal data for damage occurrence detection with the auto-associative neural network and for damage localization with the probabilistic neural network are presented. Five neural network based novelty detectors using only natural frequencies of the intact and damaged structure are first developed for the detection of damage occurrence in the Tsing Ma Bridge. The noisy/uncertain measurement data are produced by polluting the analytical natural frequencies with random noise. Numerical simulations of a series of damage scenarios show that when the maximum frequency change caused by damage exceeds a certain threshold, the occurrence of damage can be unambiguously flagged with the novelty detectors. A probabilistic neural network using noisy modal data (natural frequencies and incomplete modal vectors) is then constructed for the localization of damage occurring at the Tsing Ma Bridge deck. The main-span deck is divided into 16 segments and the damage in each segment is defined as a pattern class. The analytical modal data for each pattern class are artificially corrupted with random noise and then used as training samples to establish a three-layer probabilistic neural network for damage localization. A preliminary investigation shows that the damage to deck members can be localized only when the level of the corrupted noise is low.