MODELING UNKNOWN STRUCTURAL SYSTEMS THROUGH THE USE OF NEURAL NETWORKS

System identification refers to any systematic way of deriving or impr oving models for dynamic systems through the use of experimental data. It is an area of considerable importance in structural engineering wh ich has been gaining increasing attention over the last decade or so. Some representative publications on the subject are available in the w ork of Beck,(1) Ibanez,(2) Masri and Caughey,(3) Natke,(4) Masri and W erner(5) and the IMAC Proceedings.(6) The methods of system identifica tion provide a means of utilizing laboratory and field testing to impr ove dynamic modelling capabilities for civil infrastructure systems su ch as high-rise buildings, bridges and dams. For example, by systemati cally utilizing dynamic test data from a structure, rather than relyin g on theory alone, models can be derived which provide more accurate r esponse predictions for dynamic loads on the structure which are produ ced by wind or earthquakes. Another application is to continually upda te the model through vibration monitoring of the structure to provide a convenient method for defect identification or damage assessment.(7- 9) The potential for using active control approaches to reduce the res ponse of large civil structures under arbitrary dynamic environments, such as earthquakes, has drawn a considerable amount of interest world wide. Among the key research topics in this area is the development of system identification approaches that can cope with the challenging n ature of physical structures encountered in the structural mechanics a nd earthquake engineering fields,(10) Since the model structure in man y practical dynamics problems is by no means clear, an increasing amou nt of attention is being devoted to non-parametric identification meth ods. These methods do not identify the physical parameters of the syst em (such as mass, stiffness, etc.) but instead identify the parameters of a mathematical model which Ms the input/output data. The present p aper introduces a new non-parametric identification method for unknown dynamic systems undergoing arbitrary earthquake-type excitation. The method is based on the use of artificial neural networks as system ide ntifiers. Artificial neural networks are the ideal choice in cases whe n real time processing of large amounts of data is required because of their inherent massive parallelism, fault tolerance and learning capa bilities. In some very recent publications neural nets are used for th e detection of structural damage,(11) and for the identification of si ngle-degree-of-freedom structural systems with linear or non-linear re storing force characteristics.(12-15) In Chassiakos and Masri,(16) a m ulti-degree-of-freedom system has been identified and subsequently val idated under stochastic earthquake-like excitation. This is the type o f system studied in the present paper. Moreover, other issues presente d and discussed in this paper are: the network size and topology, the network training algorithms, validation of the identified model and it s prediction capabilities.