Machine learning approaches to determining truck type from bridge loading response

The paper is concerned with the development and comparison of alternative machine learning methods of determining the type of truck crossing a bridge from the dynamic response it induces within the bridge structure, the so-called weigh-in-motion problem. Weigh-in-motion is a rich engineering problem presenting many challenges for current machine learning technologies, and for this reason is proposed as a benchmark for guiding and assessing advances in the application of this field of artificial intelligence. A review is first provided of existing methods of determining truck types and loading attributes using both machine learning and heuristic search techniques. The most promising approach to date, that of artificial neural networks, is then compared to support vector machines in a comprehensive study considering a range of configurations of both modeling techniques. A local scatter point smoothing schema is adopted as a means of selecting an optimal set of design parameters for each model type. Three main model formats are considered: (i) a monolithic model structure with a one-versus-all truck type classification strategy; (ii) an array of sub-models each dedicated to one truck type with a one-versus-all classification strategy; and (iii) an array of sub-models each dedicated to selecting between pairs of trucks in a one-versus-one classification strategy. Overall, the formats that used an array of sub-models performed best at truck classification, with the support vector machines having a slight edge over the artificial neural networks. The paper concludes with some suggestions for extending the work to a broader scope of problems.