Identification of defects in underground structures using machine learning aided distributed fiber optic sensing

Abstract

Despite the extensive use of distributed fiber optic sensing (DFOS) in monitoring underground structures, its potential in detecting structural anomalies, such as cracks and cavities, is still not fully understood. To contribute to the identification of defects in underground structures, this study conducted a four-point bending test of a reinforced concrete (RC) beam and uniaxial loading tests of an RC specimen with local cavities. The experimental results revealed the disparity in DFOS strain spike profiles between these two structural anomalies. The effectiveness of DFOS in the quantification of crack opening displacement (COD) was also demonstrated, even in cases where perfect bonding was not achievable between the cable and structures. In addition, DFOS strain spikes observed in two diaphragm wall panels of a twin circular shaft were also reported. The most probable cause of those spikes was identified as the mechanical behavior associated with local concrete contamination. With the utilization of the strain profiles obtained from laboratory tests and field monitoring, three types of multi-classifiers, based on support vector machine (SVM), random forest (RF), and backpropagation neural network (BP), were employed to classify strain profiles, including crack-induced spikes, non-crack-induced spikes, and non-spike strain profiles. Among these classifiers, the SVM-based classifier exhibited superior performance in terms of accuracy and model robustness. This finding suggests that the SVM-based classifier holds promise as a potential solution for the automatic detection and classification of defects in underground structures during long-term monitoring.