Application of Artificial Neural Networks to Ultrasonic Pulse Echo System for Detecting Microcracks in Concrete

Abstract

Abstract : Concrete deterioration standards containing various levels of microcracks were engineered by adding calcium sulfate to the concrete mixture and curing under moisture at 38 deg. C (100 deg. F). The level of the mircocracks was classified according to the speed of the ultrasonic pulse velocity (UPV) through the specimens using the American Society for Testing and Material (ASTM) C 597 (ASTM 1994c) test method and was found to vary uniformly from 1,737 to 4,877 m/sec (5,700 to 16,000 ft/sec). After receiving some preprocessing, 186 ultrasonic pulse echo (UPE) signals were used as input training examples for the artificial neural network (ANN) system. Target values for the ANN were the measured UPVs as determined form the ASTM C 597 test method. The correlation coefficient from a least squares fit was 98.6 percent. After training, the ANN was performance tested with 30 UPE signals that the model had not seen in training. A least squares fit demonstrated that the output velocities from the ANN correlated well with the measured (target) UPVs. The correlation coefficient was 84.8 percent. The system was able to rank all six specimens in the correct order of deterioration. This investigation demonstrates that the automated interpretation of UPE signals for continuous interfaces by the ANN is feasible.