Safety evaluation of truss bridges using continuous Bayesian networks

Truss structures have been widely used in railway and highway bridges. However, it is difficult to evaluate the safety states of truss bridges due to a large number of members, high system redundancy, and insufficient monitoring data caused by limited sensors. Due to this, a new method based on continuous Bayesian Networks (BNs) is proposed to perform the safety evaluation on truss bridges, where the external loads and member stresses are assigned with two types of nodes in the BN. First, the BN topologies are established by combining the mechanical relationship of truss members to avoid the negative influence of redundant and missing edges in large-scale topology learning. Then, since the monitoring data in real-world civil structures are usually insufficient, a completed numerical sample set is adopted for parameter learning to obtain the conditional probability distributions. Subsequently, multiple BN models corresponding to different external load levels are established. A monitoring data-based matching approach is presented for model screening to find the optimal BN model. After that, the data from a few truss members are used to simulate the insufficient data in actual situations, and they are input into the optimal BN model to infer the external loads and the stresses of the remaining members. The inference ability using insufficient information well accords with the demand in engineering practice. Furthermore, two safety indices representing system states and failure probabilities are defined to evaluate the safety states of truss bridges. Finally, the feasibility of the proposed method has been successfully verified against both numerical and experimental truss bridge models.