Detection and classification of faults using maximum likelihood and Bayesian approach

Neural networks formulated in the maximum-likelihood framework and trained using the scaled-conjugate-gradient method are compared to the networks formulated in the Bayesian framework and trained using the hybrid Monte Carlo method. To make this comparison, vibration data from a population of 20 steel cylinders with various fault cases introduced are measured. To simulate faults, holes of 10-15 mm diameter are introduced in each of the cylinders. From the measured vibration data the modal properties and pseudo modal energies are extracted. Neural networks are trained using the modal properties and pseudo modal energies and are used to detect the presence of faults in cylinders and to classify the identity of those faults. The Bayesian approach is found to give the identity of faults and the respective confidence levels. However, Bayesian networks are found to be computationally expensive to train than the maximum likelihood method.