Data-driven Discovery of Real-time Road Compaction Parameters

Abstract: Road compaction is the last and important stage in road construction. Both under-compaction and over-compaction are inappropriate and may lead to road failures. Intelligent compactors has enabled data gathering and edge computing functionalities, which introduces possibilities in data-driven compaction control. Compaction physical processes are complex and are material-dependent. In the road construction industry, material physical models, together with boundary conditions, can be used for modeling effects of compacting the underlying subgrade materials and the pavement (the most widely used is asphalt) itself on site, which can be computed using Finite Element (FE) methods. However, parametrizations of these physical models require large efforts, creating difficulties in using these models to optimize real-time compaction. Our research has, for the first time, bridged the gap between data-driven compaction control and physics by introducing the parameter identification pipeline. Two use cases are investigated, corresponding to offline learning and online learning of parameters. In offline learning, a sequence of actions is learned to maximally reduce parameters uncertainties without observing responses; in online learning, the decisions of actions are made and parameters are derived while sequential observations come in. The parameter identification pipeline developed in this thesis involves compaction simulation using a simple physical model, surrogate model development using Artificial Neural Network (ANN), and online/offline optimization procedure with Approximate Bayesian Computation (ABC). The developed procedure can successfully identify the parameters with low uncertainty for the case that the selected experiments supply enough information to theoretically identify the parameters. For the case of that parameters cannot be theoretically identified by certain experiments, the identified parameters have larger uncertainties.