Fusion of heterogeneous data in non-destructive testing and structural health monitoring using Echo State Networks

Abstract

Failure to monitor the condition of key infrastructure such as roads and bridges can result in costly closures, but the economic impact could be lessened by early intervention. Non-destructive testing (NDT) examines structures without causing damage, while structural health monitoring (SHM) monitors a structure throughout its life. This thesis presents a machine learning approach to fusing heterogeneous sensor modalities that can be systematically applied to improve sensor interpretation and reduce reliance on expertise. For the first time, echo state networks (ESNs) were used in two separate NDT and SHM data fusion case studies. The NDT-based study looked at detecting defects in steel reinforcement, teaching ESNs to combine magnetic flux leakage (MFL) and cover depth data in order to compensate for variation in MFL amplitude with increasing cover depth. Using seven different cover depths between 42.5 mm and 289 mm, the fusion approach offered improved performance for 42.5mm < depth < 205mm and the most consistent calculated optimal output threshold, demonstrating the ease of systematic application. In the SHM-based study, data from the National Physical Laboratory (NPL) footbridge monitoring project was processed by a suite of ESNs to detect, localise, classify and assess damage caused by deliberate interventions. A novel approach of combining physical and environmental sensors in order to model a different modality of physical sensor made it possible to use the residual to observe damage trends and locations, which also led to the isolation of a faulty strain gauge. There was additional success in distinguishing between different intervention types and producing a metric to express the damage level. Across both studies, the ESN approach to heterogeneous data fusion improved upon non-fusion-based alternatives. This suggests that future work should consider structures that are in regular use, combining further sensor modalities and the development of bespoke data fusion software.