Disc motion control on a square aluminum Chladni plate using trained neural network

Abstract

We successfully trained a three-layer neural network (NN) to predict the frequencies needed to excite a square aluminum Chladni plate to move a disc on top of it to a specified direction. The training set includes x-positions, y-positions, magnitude, and angle of the displacement vectors of 4 resonant and 4 non-resonant frequencies. The three layers of the NN contain 32, 64, and 32 neurons respectively, and was trained using a tanh activation function. The results of the trained model show a validation and test accuracy of 79.4% and 81.5%, respectively. We found that the errors are from frequencies that have nearly the same displacement vectors at certain positions as evidenced by the overlay of displacement vectors of the predicted frequency versus those of the desired frequency. Because they have the same direction, the misclassified frequencies still contributed to moving the disc to the desired location. A disc was successfully moved 69 out of 80 times to desired locations where the 11 failed attempts were due to the disc falling off the plate prematurely.