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

136

6

Fig. 6. (a) Laboratory specimen of GFRP plate with piezoelectric transducers and damages artificially introduced; (b) Scheme of the signal paths in case of the first transducer served as the actuator of elastic wave.

The second damage was the impact and it was located in the upper right quadrant of the sheet (see Fig. 6a). For this case of damage another ten series of measurements were stored. Hence, ultimately there were 20 measurement series for the undamaged specimen and 20 measurement series for the damaged specimen available. It can be readily appreciated that the main difficulty in analysis of the signals stored in the patterns database is the significantly larger amount of data, when compared with the one-dimensional strip described in Section 3.1. Nevertheless, the approach used previously for the GFRP strip can also be used for this particular specimen under test. Let us assume for this purpose, that each path from an actuator to a sensor is considered as a signal propagating in a strip specimen, eg. path 1-5 in Fig. 6b. Then, signals measured on this path and stored in the patterns database can be processed by PCA. If we assume that an input vector for a single ANN contains 16 principal components of each signal (which represented in this case 99% of information), this vector has 16 rows and 40 columns. After successful training of the ANN for novelty detection, for each of the signals the obtained values of NI can be assigned as well. In this way, the multidimensional data were reduced and then the NI values calculated for all patterns can be collected in a single matrix of 8x8x40. This data already contains information about the occurrence of damage but this knowledge is based only on the signals received by single sensors. One can imagine that in case of damage each of the individual receivers at their position may turn on the red light. Obviously the emergence of damage can be based on measurements from a single sensor, but inference can also rely on a network of all sensors because the further reduction of data size can be achieved through the calculation of the principal components form the NI data. In this way, in each case of a single actuator it was possible to determine one parameter representing the condition of all receivers in the current test. In other words, this parameter should determine whether a signal actuation from a particular transmitter will indicate the possibility of a damage occurrence at any receivers. Results of the thereby performed classification led to explicit detection of damage. Again, the NI values do not allow a clear distinction of two different states of the damaged structure (D1 relates to a single damage, while there are two cases of damage in the case study D2). Therefore, at the second level of the diagnostic system a NN for regression task was designed. Its aim was to identify the considered damage scenarios. For this purpose each of the individual patterns has been assigned with the following labels: ''0'' for undamaged state, ''1'' for the first damage (chemical corrosion), ''2'' for state with two kinds of damages (chemical corrosion and impact). The input vector was defined by eight values of NI taken from each of the sensors (more precisely from the output vector of the respective ANNs trained at the first level). Although half of the pattern was used for testing, good identification results of particular structural states and damage types can be noticed. 3.3. CFRP plate The last example is related to anomaly detection in a CFRP plate. This study case differs from the previous one mainly by varying locations of introducing discontinuity. Although it would be interesting to identify the exact location of discontinuity, only some preliminary results of that identification were presented in this chapter.