Evolutionary safety evaluation of a truss bridge using dynamic Bayesian networks

Timely safety evaluation of a real-world complex truss structure is difficult because of complex failure modes, uncertainty influence and difficulty in theoretical deduction. Therefore, a dynamic Bayesian network (DBN) has been designed to demonstrate the safety evolution process of a truss bridge under a time-varying load. The DBN comprises a prior network and a transition network, forming different time slices. Its network nodes represent the truss members and system, and the discrete nodal variables indicate the probabilities for safety and failure states. An effective network topology definition method is proposed by incorporating a hybrid topology learning strategy with a virtual substructure division strategy. The two strategies provide a rational topology with the reduced dimensions of conditional probability tables for complex truss structures. Numerical observation data are generated for learning the conditional probabilities between connected nodes in both the prior and transition networks. Subsequently, state probability inference between different time slices can be achieved using measured observation data from a limited number of members at a given time as the evidence. Afterwards, the failure state probability evolution curve of the truss bridge system can be described. The validation on an experimental truss bridge model has successfully demonstrated its state evolution under the different loading periods. The failure time of the truss system was predicted, which well accorded with the experimental observations.