Peeter Akerberg Ben H. Jansen
Dept. of Elec. Eng., Univ. of Houston, Houston, TX 77204-4793
Shail R. Pandya Zhijing Wang Robert D. Finch
Univ. of Houston, Houston, TX 77204-4792
Artificial neural networks (ANNs) are being used for the detection of cracks in metal and concrete beams and structures. Vibration signals are obtained from these structure by impacting with a small hammer and recording the activity of an accelerometer attached to the structure. These vibration signals are digitized, and the samples are input to an ANN. Detection of cracks proceeds in two different ways. For both methods, the networks are trained on data obtained from intact structures, but in the case of the first method, training is halted once the ANN can accurately predict future samples of the vibration signal from present and past observations. These ANNs are then fed with data obtained from structures with varying degrees of defects, and the prediction error is noted. In the second method, training continues after a small defect has been made to the structure, and the weights of the nets are compared. With both methods, we can detect cracks as small as 0.1 in. Examples of the results obtained for metal and concrete beams will be shown. [Work supported by NSF Grant no. MSS-9024224.]