<title>Location and magnitude of impact detection in composite plates using neural networks</title>

Abstract

A method of determining the location and extent of impact induced damage in isotropic plates is investigated numerically and experimentally. A computer simulation of impact to this model plate provided information in the form of frequency response data and strain traces. These data sets were inputs into two different backpropagation neural networks. The first neural network uses FFT of strain data, which is split into real and imaginary parts and integrated over the first ten harmonic natural frequencies for each of four sensors (8 inputs into neural network). This particular network has been trained on a set of 63 simulated impacts, with the network solving for the location of impact sites within an average of 8 millimeters per impact. The second network uses the same FFT of strain data (8 inputs), but additionally uses the impact coordinates as inputs (total of 10 inputs). The network has been trained on a set of 800 simulated impacts, and solves for magnitude of the impact events with an average error of 7.08%. Experimental impact location determination has been accomplished by using a set of four strain gage sensors mounted in the corners of the isotropic aluminum plate. Data from these sensors were post processed as mentioned above, and used to train a backpropagation neural network. A twenty impact subset of a 72 impact grid was trained to locate impacts with an average RMS error of 1.019 radial centimeters.