Feature extraction of GPR data for the classification of reflectors in reinforced concrete

More Info
expand_more

Abstract

During their service live reinforced and prestressed concrete structures experience deterioration. Therefore, to effectively assure the safety and durability of the compounds, non destructive testing methods, such as Ground Penetrating Radar, employed at high frequencies, provide a rapid assessment of the geometry of the compound. Although amplitude and traveltime evaluations are the most widely used techniques, attribute based analysis to enhance the interpretation quality has gained interest in recent years. A great quantity of surveys exist that attempt to determine the diameter of cylindrical targets such as reinforcement bars or tendon ducts. Indeed, the affiliation of a reflector to one of the two groups is still a challenging task. Therefore, different attribute approaches are applied to four data sets with high spatial sampling rates collected under controlled conditions, to derive features that unambiguously assign the considered reflection hyperbolas to either reinforcement bars or tendon ducts. For this purpose, B-scans are reduced to the individual reflectors by manual selecting the desired windows. Then, attributes are applied to the hyperbolas and the results are compared with each other to derive features describing the two reflector types. For attributes where features can be extracted, they are tested for measurements with relatively low spatial sampling rates and for one on-site data set. It is shown that the analysis provides sufficient classifications for six out of the nine tested attributes when data with high spatial sampling rates under controlled environment conditions are considered. In contrast, relatively low spatial sampling rates yield to accurate results for only three attributes: The maximum absolute amplitude, the cumulative energy and the amplitude at the apex. Although promising results for low spatial sampling rates are obtained, the application to on-site data is still complicated. Nevertheless, this study provides the basis for automated reflector classification.