Current issues in vibration-based fault diagnostics and prognostics are: (a) fault/damage identification; (b) localization; and (c) monitoring. The fault/damage identification and localization are needed for diagnostics. The monitoring allows observing the fault/damage progression in time. Based on monitoring, the prognostics of the moment of failure, and hence of the remaining life could be performed. Critical in both the diagnostics and the prognostics is the existence of a reliable damage metric. Several methods for achieving this, from single parameter classification, through overall-statistics classification of the entire spectrum, to probabilistic neural network (PNN) method are presented. These methods are illustrated on: (i) simulated data, (ii) experimental data taken on simple- geometry calibration specimens; and (c) experimental data recorded on aging aircraft panels. The experimental data was obtained with the electro-mechanical (E/M) impedance technique using small and unobtrusive piezoelectric-wafer active sensors (PWAS) bonded to the structure. This experimental data was in the form of very high frequency (hundreds of kHz) spectrum representing the structural dynamics in sensor's neighborhood. The overall statistics approach, which is simplest to apply, could correctly classify near-field spectra using the dereverberated response and the correlation coefficient difference (CCD) damage metric, but it could not classify medium field spectra. The features-based PNN could correctly identify both medium-field and near-field data, but it requires more elaborate, two-stage, data processing that involves first features extraction and then classification.