Structural condition assessment of the Bill Emerson Memorial cable-stayed bridge using neural networks

Structural health monitoring is a challenging task that has recently received great attention from research communities. Due to its ability for predication and classification, neural network has become a promising tool for the identification of structural damages. In this dissertation, two backpropagation neural networks are introduced to predict the structural responses and to quantify the structural damages. The proposed methodology differs from many existing technologies in that it can be used to detect damages directly from the measured time signals without requiring modal characteristics. The methodology is developed with a benchmark highway bridge and is implemented for the Bill Emerson Memorial Cable-stayed Bridge. Extensive analyses indicated that the performance of an emulator neural network for response prediction is independent of its training sets after a novel prediction error indicator, response weighted root-mean-square (RW-RMS) was introduced. In comparison with the RMS error, the proposed RW-RMS is more suitable for damage detection since damage is typically associated with the peak response of a structure. As validated with the highway and the cable-stayed bridges, the parameter evaluator neural network can effectively quantify damages as small as a 10 percent reduction in flexural stiffness. To simulate the "healthy" structure and introduce various damage scenarios, a 3-dimensional FEM of the cable-stayed bridge is established and validated with the measured earthquake data. The first 22 natural frequencies of the bridge model agree with the measured frequencies by a less than 8 percent difference.