A problem of current interest is the automatic detection and classification of faults in mechanical systems such as the gearboxes and transmissions on board helicopters. Current fault processing uses relatively simple metrics to characterize changes in measured vibration data. The metric specification and thresholds for detection and classification are found using complicated analytic models for the gearboxes. With this approach finding a solution to the problem is difficult since the understanding of the problem, the metrics that can be used, and the fault classes that are accounted for are only as good as the model developed. In many cases, the interaction of fault conditions with the mechanical system is time varying and highly nonlinear. An alternative solution described here is to use generalized time-frequency and time-scale representations coupled with a hierarchy of neural nets. The processing assumes no underlying model for the events of interest. Rather the system 'learns' to detect and classify faults by examination and fusion of features from training data which have known fault conditions. Results of processing real helicopter gearbox vibration data with seeded faults are shown.