PSI - Issue 5

Piotr Nazarko et al. / Procedia Structural Integrity 5 (2017) 131–138 P. Nazarko et al./ Structural Integrity Procedia 00 (2017) 000 – 000

135

5

regression tasks. The principal components calculated at the first stage may now be used to assess the damage parameters. In our example the NI values do not allow of clearly separating the two last classes of damage (see D3 and D4 in Fig. 4.), so a neural network 16-3-1 was designed and trained in order to determine the appropriate level of the damage. An output vector consisted of a real number ranging from 0 to 4, which corresponds to an undamaged specimen as well as the four damage levels. The label ''0'' was introduced in case the novelty detection signal gives a false alarm (this may happen at low threshold levels, but this scenario is safer for a monitored structure than misclassified damage). Exemplary results of the damage level identification are shown in Fig. 5.

Fig. 5. Damage level identification by NN (16-3-1): (a) Learning results; (b) Testing results.

3.2. GFRP plate

In the next stage of the study the laboratory experiment was carried out on a sample of GFRP composite reinforced by fibreglass roving arranged in four layers ±45°. Dimensions of the plate studied were 200 x 312 mm, while its total thickness was about 1.8 mm. On the surface of the analysed specimen eight piezoelectric transducers (Noliac CMAP03) were mounted using epoxy adhesive (Fig. 6a.). Thus, when one of them was used to force the wave propagation, the other were employed as sensors and structural response was stored by the acquisition system used (PAQ-16000D). As the extortion pattern a package of four sine waves (with frequency of 320 kHz) was adjusted and modulated by Hanning window. In the case of the analysed sample of GFRP sheet twenty series of measurements was related to intact specimen (undamaged). Each series consisted of seven time signals recorded by the receivers for the successively changed actuators. These seven signals are related to the paths shown schematically in Fig. 6b. The row of the excitation signal was filled with zeros. As a result, after each series a matrix 8x8x12500 was created, where the third dimension was related to the length of the time signals. Thus, in the case of the specimen without damage there were twenty of such matrices (repetitions) in the patterns database. The first damage was introduced in the form of chemical corrosion. Local acid action (48 hours) caused damage to the composite epoxy matrix. The resulting die pocket was 12 mm in diameter and has a 0.5 mm depth (located in the middle of the left half of the sheet, Fig. 6a). For this case of damage ten series of measurements were stored.