Deep Learning Framework for Total Stress Detection of Steel Components

Abstract

Accurately obtaining the stress of steel components is of great importance for the condition assessment of civil structures. Currently, methods for structure stress detection have some drawbacks such as the capability of obtaining structural stress increment rather than the total stress, causing structural damage, and high cost. To overcome these drawbacks, a deep learning framework for total stress detection of steel components is proposed and its feasibility is illustrated with an example. First, the adopted deep neural network is briefly introduced, followed by the introduction of the dataset preparation. In order to maximize the stress detection accuracy, parameter analysis was conducted and the mean average precision achieved by the well-trained model for detection of the stresses under consideration is 89.67%. The robustness of the trained model was further examined and the procedures for application of the proposed approach were summarized. The presented method provides a new idea to detect the total stress of structure components that is difficult to obtain with a traditional sensor-based method.