ABSTRACT
The connectivity reliability analysis of bridge networks has been the focus of research in this field and is also a difficult problem because of NP-hard problem. In particular, when performing network strategy or optimization, it is very time-consuming as the network connectivity probabilities need to be solved hundreds of times. This study finds that neural network with fewer neurons and layers can quickly and efficiently solve the all-terminal connectivity reliability for large-scale bridge networks. Firstly, for the same network benchmark model, based on real bridge inspection data, the state probability samples of each edge in the network are extracted by data fitting, and Monte Carlo simulation is applied to approximate the all-terminal connectivity reliability of each network model under each sample. Secondly, multiple neural network models containing one, two and three layers with different neurons are proposed and 5-fold cross-validation is applied to obtain the hyperparameters of each model. Finally, in conjunction with appropriate optimizers and hyperparameters, it can be shown that simple single-layer network models can find sufficiently accurate all-terminal connectivity reliability for large-scale bridge network. The accuracy and efficiency of the pro-posed method is verified based on a real bridge network containing 216 directed edges with 63 nodes.
