A new methodology to diagnose pavement subsurface condition using ground penetrating radar

This thesis presents a new methodology to assess pavement subsurface condition using Ground Penetrating Radar (GPR). The methodology is based on an Adaptive Neuro-Fuzzy Inference System (ANFIS). Time series and wavelet analyses of GPR signals are employed along with a trained ANFIS to estimate an aggregated rating of pavement subsurface condition. The developed methodology provides an effective means that mimics subsurface evaluation of pavements by human experts.; The ANFIS consists of a number of rules that map CPR signal parameters (input) into a pavement health condition rating (output). These rules are adaptively modified during training of the fuzzy system to optimally refine the mapping of the input into output data. A simple Sugeno fuzzy model with eight rules is used in this study. Reference signals corresponding to intact and deteriorated pavement sections are used in the training process. The training data is obtained from visual evaluation by human experts of reference pavement sections. Time series and wavelet analyses of pavement GPR data are employed to obtain features and information on layer interfaces which are then used as input parameters in the ANFIS rating analysis. These inputs parameters were selected to consist of maximum amplitude of raw GPR signal, mean absolute deviation (MAD) of raw GPR signal, cross-correlation of the GPR signal with reference signals, cross-correlations of wavelet subband histories of the CPR signal, magnitude of approximate coefficients, and MAD of wavelet approximate coefficients. These different input parameters are processed using the trained ANFIS to obtain an aggregated linguistic pavement rating consisting of good, moderate, and deteriorated conditions. For a given pavement section, the outcome of the ANFIS analysis is presented in terms of contour maps, and receiver-operating characteristic (ROC) curves are used to validate the model. A case study was analyzed and showed less than four percent of "miss" and "false alarm" results, with about 80% "hit" rate of overall pavement condition rating. Thus, the model may be considered adequate in practical applications, especially in view of the high noise level naturally associated with human evaluation.