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

137

7

Laboratory measurements were performed on a CFRP plate specimen, which was equipped with a grid of 12 permanently mounted piezoelectric transducers Fig. 7. Similarly to the previous tests, the excitation signal (four sine waves of 60 kHz modulated by Hanning window) and specimen responses were generated and recorded by PAQ 16000D system (Fig. 2b). Each time one of the transducers served as the actuator, while the other 11 transducers as the sensors of elastic wave signals. One series of measurement consisted of 11 x 12 signals. A discontinuity in form of stress field concentration (hand screw) was introduced into the laboratory specimen. Measurements were preformed for 24 undamaged states and 35 various localizations of discontinuity. These anomaly locations (AL) were marked in Fig. 7. together with paths of Lamb waves propagation.

Fig. 7. Investigated area of the CFRP plate with piezoelectric transducers (PZT), anomalies artificially introduced (Anomaly location, AL) and scheme of the signal paths.

The first objective of the proposed algorithm is detection of anomalies in form of the introduced discontinuities. In this approach each of the signals has been processed to the domain of principal components. Then, at the level of each receiver, ANNs (16-3-16) were trained for novelty detection. The information collected from all the sensors enabled clear detection of appearing anomalies. At the next step of anomaly detection an attempt was made to identify its location. A NN for regression task was built for this purpose. The input vector consisted of NI collected from all the sensors, while the output vector consisted of two rows filled with the anomalies coordinates (x i and y i ). Results of this preliminary test are collected in Table 1, in the form of prediction errors. There are mean values of mean squared error (MSE) and correlation factor (R) calculated from 50 trained NNs in Table 1.

Table 1. Identification errors of anomaly positions.

Learn

Test

Coordinate

MSE

R

MSE

R

x i y i

0.808 ± 0.032 0.889 ± 0.026

0.90 ± 0.02 0.95 ± 0.01

0.764 ± 0.032 0.887 ± 0.025

0.764 ± 0.03 0.89 ± 0.03

4. Conclusion

In the paper the operating concept of a two-stage diagnostics system was presented. Its main task was a novelty detection and an assessment of defects detected. The process of inference was based on changes in the recorded signals of elastic waves and NNs trained for anomaly detection and prediction of its parameters. The PCA has significantly reduced the problem of processing multidimensional data and enabled the correct training of neural networks. It should also be noted that in case of data recorded in the GFRP strip specimen, patterns