Recurrent Neural Networks Considering Natural Conditions for Determination of the Reinforcement Degrees when Constructing Infrastructure Structures

Abstract

In recent years, the use of AI has been promoted in the construction industry to solve labour shortages. This study focuses on determining the reinforcement degree when constructing a kind of infrastructure, which has required the empirical knowledge of experts, and proposes a machine learning method to automate the process. The target structure is divided into many sections, and the reinforcement degree for each section needs to be determined based on observation data.

In this paper, we formulate the problem as a time series recognition one in which the degree is determined using observation data from the current construction section and previous sections already constructed. Then, we apply a recurrent neural network (RNN) to the problem, which has a feedback mechanism suitable for time series processing. In addition, we propose RNNs which use different weights corresponding to observation data at different construction site. Through verification experiments using datasets observed at real construction sites, we show the advantages of the proposed RNNs by comparing the existing methods.