Generalized method for distributed detection and quantification of cracks in bridges

Abstract

The development of a generalized machine learning approach based on distributed detection and quantification of cracks by optical fibers is described in this article. A Brillouin scattering optical fiber sensor system was employed to develop, test, and verify the method. The main components of the approach described herein consist of an unsupervised crack identification module based on the iForest algorithm and a crack quantification component by the one-dimensional convolutional neural network method. The main attribute of this model is the versatility for application in various types of structures. The proposed method does not require further application-dependent training or calibration as long as the structural applications employ the same optical fiber type and installation adhesives. The effectiveness of the proposed method was verified by two experiments involving a 15-m steel beam in the laboratory and monitoring a twin set of 332-m-long, five-span continuous box girder concrete bridges. Regarding crack detection capabilities, it was possible to detect 107 out of 112 cracks in the laboratory beam and 20 out of the 21 in the bridges. The resolution of crack opening displacements for the steel beam and concrete bridges were 20.6 and 21.7 µm, respectively. The verification experiments further indicated the generality of the approach in applications to various types of structures and materials.