Fault classification using pseudomodal energies and neural networks

A new fault identification method is introduced that uses pseudomodal energies to train neural networks. The proposed procedure is tested on a simulated cantilevered beam and a population of 20 cylindrical shells, and its performance is compared to that of the procedure that uses modal properties to train neural networks. Both the cantilevered beam and cylindrical shells are divided into three substructures, and faults are introduced into these substructures. The cylinder is excited using modal hammer, and acceleration is measured using an accelerometer. Each fault case is assigned a fault identity with the presence of fault represented by a 1, whereas the absence of fault is represented by a 0. Following this fault representation scheme, a fault located in substructure 1 would have an identity of [1 0 0], with two zeros indicating the absence of faults in substructures 2 and 3. The neural network used is a multilayer perceptron trained using scaled conjugate method. The statistical overlap factor and principal component analysis are used to reduce the size of the input data. For both examples the pseudomodal-energy-trained neural networks provide better classification of faults than the networks trained using the conventional modal properties.