Summary
A neural network is designed and trained using a supervised learning approach to localize damages on plates. Mindlin's theory is used to derive the plate's equation of motion, and the finite element method is used to discretize it. Damages are simulated by reducing the thickness (or stiffness) of the plate in the damaged area. The equation of motion is solved in the time domain for various locations of damage and different external forces for generating the training data for the supervised learning method. Transverse displacements are stored for multiple instants of time along the entire plate's surface. The data is converted into a one-dimensional vector and utilized as training data. Neural networks are demonstrated to be able to localize damages well even for excitation frequencies that are not utilized in the training set.