Fault Detection in Steel-Reinforced Concrete Using Echo State Networks

Abstract

The cost of maintaining and repairing the world's ageing reinforced concrete infrastructure continues to increase, and is expected to cost the United States economy alone $58 billion by 2020. Consequently, the use of non-destructive testing technologies for the early identification of faults in roads and bridges is becoming increasingly important. One such technology is the Electromagnetic Anomaly Detection (EMAD) technique, which exploits non-destructive magnetic flux leakage to detect defects in steel reinforcing meshes embedded in concrete. Despite the increasing need for such techniques, the data analysis options currently in use are limited. This paper presents an application of Echo State Networks, a recurrent neural network from the field of reservoir computing that features a short-term memory, to data obtained using the EMAD technique. Having been trained to discern real defect signals from other anomalous magnetic features, the performance of the ESNs was then compared to that of an analytical data analysis technique that is currently used to process EMAD data. It was found that average ESN performance was comparable in terms of AUC, while the optimal threshold was more consistent, greatly aiding application in the `real-world'. A qualitative analysis of the output of both methods on an unseen testing dataset also demonstrated the superiority of ESNs for practical use as a real time tool for on-site inspections.