a-has any structural damage occurred?

b-where did it occur?

c-how severe is it?

Any crack or localized damage in a structure reduces the stiffness and increases the damping in the structure. Reduction in stiffness is associated with decreases in the natural frequencies and with modification of the modes of vibration of the structure. In some cases, the damage position and magnitude can be identified from these changes.

Natural frequency measurements are a very attractive form of NDT because they can be carried out at a single point of the components and they are also relatively independent of the position chosen. Such a measurement can be completed in some seconds if a tap-testing technique is used and the test can readily be automated for use on a production line.

The direct problem is well solved, even for composite structures. Modal Analysis, Finite Element Methods and Perturbation Methods are some tools used for that purpose. Meanwhile methods for the solution of the inverse problem (determination of what has happened to the structure for a given modification in the dynamic behaviour) are not so much developed. The design of analytical methods for this purpose is much more complex, and can usually give good results for very simple structures with a restricted number of defects. But in NDT applications, typical type of defects are usually known before. So for the purpose of NDT it is "only" necessary to recognize patterns in the observed dynamic response of the structure resulting from damage, including the capability of determining the extent of damage. This capability appears to be within the scope of the pattern matching capabilities of artificial neural networks.

Artificial neural networks (ANNs) are systems composed of many simple computational elements densely connected by links with variable weights. The most current use of ANNs nowadays is in areas such as speech and image recognition where many hypotheses are pursued in parallel, high computation rates are required, and even the best systems are currently far from equalling human performance. Meanwhile, recent research is showing the possibility and benefits of applying this technology to other areas like identification of nonlinear dynamic systems and NDT. In this study, neural networks are used to extract and store the knowledge from the patterns of the response (natural frequencies) of the damaged structure at different locations.

A feedforward ANN with a back propagation learning algorithm (supervised learning) is trained to recognize damage locations (output) from the extraction of features from changes in the natural frequencies (input). In the proposed methodology, the training phase of the ANN is used to "extract" the cause and effect relationships between the natural frequencies and the damage locations. After the training phase has been completed, those relationships are stored in the connection strengths of the ANN. In this way, the self-organization and learning capability of these ANNs are utilized to eliminate the need for explicitly extracting the cause and effect relationships.

The training set of the ANN with "appropriate" data constitute the central point of this approach. The obtaining of this "appropriate" data set is normally very costly. At this work the data for the training of the ANN will be obtained numerically with the use of a finite element model. Experimental data for a very simple structure like a beam will be used to show the potential and limitations of this approach. The influence of different ways of pre-processing the data on the results will be discussed too.