Application of convolutional neural network in stress calculations of reinforced concrete slabs of road pavements

Abstract

The design of rigid reinforced concrete slabs of foundations, slabs, road surfaces is based on calculation models, which are developed on a relatively limited number of experimental studies, in most cases requiring quite large material and time costs. The complex stress-strain state occurring in stiff reinforced concrete base and pavement slabs under load, especially under cyclic dynamic loading, can often lead to cracking and failure of the slabs. In this paper, reinforced concrete slabs of a container yard pavement were investigated for load bearing from the wheels of a reach stacker (container loading vehicle) travelling on the surface. Existing models for the design of such slabs typically consider the slab loaded by a single local load applied to an edge or corner of the slab from the wheels of a moving vehicle. In fact, there may be two wheels on the slab, resulting in more unfavorable conditions. The application of the finite element method in such problems is quite laborious as it requires highly skilled design engineers and considerable time, making the design routine and of limited use. This paper investigates an alternative approach based on the application of an artificial convolutional neural network (CNN) with U-Net architecture, which provides a reasonably accurate prediction of stresses in the slab much faster and simpler compared to the finite element method. The paper presents the architecture of the neural network with an indication of the features and stages of its training. Statistical analysis of the calculation results is performed, which allowed us to assess the reliability of the neural network model for determining stresses in reinforced concrete slabs on an elastic base.