A mechanical Bayesian network was used to express the logical relationship among truss members for safety evaluation of a truss structure. The network topology was established based on mechanical analysis, and the nodal variables included the maximum stress of all the members, the state of the discrete lower longitudinal girder, and the system state. The logical strength between two relevant member nodes was described by conditional probability. A mechanical Bayesian network was established based on numerical analysis and sampling. The four maximum member stresses of the steel truss structure model were input into the established Bayesian network as the known evidence. The maximum stresses of the other members, as well as the state probability of the system, were then deduced. Results show that the mechanical Bayesian network can be used to well estimate the maximum stress of the other nodes (truss members) when the maximum stresses of the monitoring members are known. The coefficient of determination, i,e., R² = 0.9992, shows that the consistency between the evaluated stress of all elements and the observed data is strong. Simultaneously, the members closer to the known members are estimated in a more accurate way, which can be used as a reference for choosing monitoring members. Furthermore, the inferred state probability of truss system is consistent with the observed data. Thus, the failure probability can be used as an index for safety evaluation of truss structures.