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

133

3

Fig. 1. Scheme of detection and damage evaluation system.diagnostic system. In general, when a trained ANN is fed with the inputs obtained from a damage state of the system, the novelty index NI , which measures the distance between the known input and output of the ANN, will increase. In this way it is possible to signal any novelties which may occur in the patterns database. At this stage the main differences in the functioning of the system in case of 1D and 2D models is the number of PZT sensors used and ANNs to be trained. In this paper, a superior ANN was designed to classify the state of the whole investigated area. All studied NNs were trained using Levenberg-Marquardt backpropagation algorithm. Patterns of the database defined can be then used to train neural network to solve individual regression tasks, i.e. prediction of damage location, its type or extent. Although this paper focuses mainly on the task of novelty detection in composite materials, there is also an example of damage extent prediction considered in the introductory case of a strip specimen. To understand the operating principles of the proposed diagnostic system, let us first consider a strip specimen, which may be treated here as one-dimensional (1D). The measuring signals of Lamb waves were obtained during laboratory tests of Glass Fiber Reinforced Polymer (GFRP) strip with dimensions 600 x 25 x 1.4 mm. At both sides of the specimen piezoelectric transducers (Noliac CMAP03) were glued: one served as an actuator, the other as a sensor. Locations of the transducers and the laboratory equipment are shown in Fig. 2. An extortion signal consisted of four sine cycles modulated by Hanning window and its operating frequency was set to 23 kHz. First, a few series of tests were performed on the intact specimen. It enabled establishing a database of 900 patterns (time signals) related to the undamaged case. Next, the first defect was considered as damage caused by the local impact of high temperatures (240°C for approx 10 minutes) at a distance of 195 mm from the left end. The consequence was visible on the surface exposed to high temperatures (dark color and micro cracks), while on the second side of the sample there were no any visible changes in the specimen's structure. This action was repeated four times, each time increasing the area influenced by high temperatures. Consequently, for each instance of the damage case (D1-D4) 225 signals were stored, which led to the definition of 900 damaged patterns. Visual assessment of the recorded time signals is not easy, because the changes caused by damages are seemingly unimportant Fig. 3. The separation of patterns becomes much simpler if the recoded signals are subjected to Principal Component Analysis (PCA). As a result, a correct patterns separation (damaged vs. undamaged) was possible based on the projection of data on the directions of the first two principal components. 3. Novelty detection and damage evaluation 3.1. GFRP strip