Normalising flow-based vehicle load modeling for bridges

Abstract

Accurate vehicle load modeling is critical to the maintenance, reliability analysis and safety assessment of bridge structures. The weigh-in-motion (WIM) system is usually used to archive actual vehicle load by recording vehicle load variables including the number of axles, axle weight, total weight, vehicle length, inter-axle spacing, and vehicle speed. However, modeling multidimensional vehicle load variables considering their correlations remains a challenge. Normalising flow, as a probability estimation method based on deep learning, models vehicle load variables as a multivariate distribution so that not only their marginal distributions but also their correlation structure can be captured accurately. This article proposed a normalising flow-based method to model the vehicle load for bridges. The normalising flow-based vehicle load model learns from the actual vehicle load variables by maximising their log-likelihood and generates new samples of vehicle load. With the proposed vehicle load model and headway information, stochastic vehicle flow can be established. The results of the WIM data from an actual bridge show that the proposed method can model the vehicle load variables accurately and the stochastic vehicle flow based on the proposed method can accurately reflect the actual vehicle flow.