Prediction of the severity of exceeding design traffic loads on highway bridges.

Being a driver of failure consequences, forecasting the severity of events where design traffic load limits on bridges have been exceeded (DLEEs) is fundamental for road safety. Previous research has focused on estimating failure consequences by direct and indirect cost metrics. Only recently has research assessed severity unconventionally, in which the type of DLEEs was predicted by applying econometric models through Binomial Logistic Regression (BLR). Since machine learning models using Artificial Neural Networks (ANN) have not yet been explored, this study will enhance the literature as follows. First, two different 'severity' models were set up as a function of bridge-side, temporal-context, and traffic load hazard variables. Whilst the former relied on a BLR, the latter used an ANN. Second, the performance of these models was assessed using confusion matrixes, some performance indicators, and a cross-entropy parameter. Raw Weigh-In-Motion data on 7.4 M+ individual vehicle transits on a bridge along a primary roadway in Brescia (Italy) were processed. Although a similarly strong performance was achieved for BLR and ANN, the results indicated that ANN was able to predict severity records with a higher level of confidence than BLR on the case study dataset, with the cross-entropy of the ANN less than one third of that of the BLR. These analyses can support road authority traffic management to safeguard bridges from traffic load hazards. Finally, this study recommends future developments, such as considering the structural effects of traffic loads in the modelling, prioritizing traffic management actions among bridges at network level, and exploring the impact of ANN models in risk assessment.