Artificial neural networks for the interpretation of ground penetrating radar data for infrastructure condition assessment

Properly operating infrastructure, such as transportation networks and facilities, is essential to the economic health of a region. Continued assessment of the condition of such infrastructure is key to effective management, which ultimately leads to efficient operation. Unfortunately, the size and extent of transportation networks in any jurisdiction makes adequate and timely condition assessment exceedingly difficult at best. Because of this, better methods have been sought to improve the quality and timeliness for condition assessment. Ground penetrating radar (GPR) has been gaining acceptance as a tool for infrastructure condition assessment for the past 25 years. It has been proven to provide quicker and more complete infrastructure network data for several infrastructure condition attributes of interest. However, GPR assessment at the desired level of detail creates an immense amount of condition data. Recently, artificial neural networks (ANNs) have been applied to the problem of interpreting GPR condition assessment data. ANNs allow automated interpretation of condition data in real-time (as it is being collected), which eliminates the problem of timeliness.; Past approaches to the interpretation of sensor data, such as GPR condition assessment data, using ANNs have been one-step monolithic methods. These methods involve using a single neural network to map raw, or slightly processed, GPR data directly to properties of the environment, such as pavement layer thickness, or moisture content, for pavement condition data. This thesis describes and investigates a segmented approach to pavement condition assessment using ANNs. This approach follows the steps that are performed during manual interpretation. During manual collection and interpretation, the user varies radar settings until the data is of sufficient quality to recognize reflection events, and then collects the data. Event locations and amplitudes are then determined from the data, properties are calculated about the data, from these events, and from these properties decisions made are in accordance with the objectives of the investigation.; Following these steps (data quality determination, event location determination, and parameter calculation) with ANNs provides better condition assessment results than performing condition assessment in one step with a single ANN. The decision making step adds another level of evaluation not contained in other automated methods. For the investigation described, the segmented approach determined asphalt layer thickness values approximately two times more accurately than the one-step approach performed on the same data.