The problems involved in structural damage identification by means of pattern recognition neural techniques are addressed. As is known, mechanical system recognition can be achieved by making appropriate use of these connectionistic instruments. Recognition takes place on the basis of an incomplete set of data contained in the system’s dynamic response.
Some methods, previously developed by the authors on the basis of numerical models, are validated through the use of noisy data. To this end, an experimental research was carried out on simply supported beams. The tests made it possible to acquire a wealth of dynamic response data from beams damaged artificially to produce defects of varying entity and position. The data obtained, appropriately normalised, are used as inputs for supervised neural networks. In particular, frequency analysis data are able to provide a summary characterisation of the distortion in a structure’s dynamic behaviour.
Different signal processing and analysis procedures are compared in order to identify the best resolution and sensitivity capabilities in the diagnostic performance of neural networks.