The subject of the paper in here is concerning of the same idea, namely the one of recovering the elastic constants from the ultrasonic signals generated through an isotropic viscoelastic composite material and the use of the recorded signals to classify the materials by mean of neural networks.

The samples used for the elastic constants recovering were 3 epoxy plates, having the same dimensions and shape (8 mm in thickness and 100 mm in diameter), but different values of density because of their powder concentration in the compositional structure. The measurements were carried out using a pulsed Nd: YAG laser, operating at 1.064 micro-m, in the ablation regime. A cylindrical lens was used to focus the generation beam to a line, 40 mm in length and 0.1 mm in width, respectively, on the sample's surface.

The ultrasonic disturbance was detected on the other side of the sample by an optical heterodyne probe. After establishing the epicentral position, signals were recorded for laser pulse duration of 20 ns, having the pulse energy of 20 mJ, working in a scanning fashion. The distance between the scanning lines was 0.25 mm, in only one side from the normal position. The total length that was scanned ranged up to 22 mm, the acquisition step being a time consuming. After fulfilling the above steps, an artificial neural network as a pattern classifier was proposed. The architecture of the network is the well-known multi-layer perceptron model, using the back-propagation algorithm for training. A three-layered network, as this case, is sufficient to generate a complex decision boundary if enough neurons are provided at the hidden layer. There was a trial and error runs to establish this number.

The outputs were 2 classes, corresponding to the types of materials that exhibit isotropic and isotropic viscoelastic behaviour.

Obs: The experimental work was carried out at the Laboratoire de Mecanique Physique, Universite de Bordeaux I, under a Socrates program.