PSI - Issue 37

Claudia Barile et al. / Procedia Structural Integrity 37 (2022) 307–313 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

308

2

which these attenuations are used to quantify the structural integrity of the material (Vary A, 1988; Barile et. al., 2019). In this technique, an artificial stress wave is generated in the material by means of external source. The propagating stress waves are recorded using a piezoelectric sensor and are analysed. In this research work, the structural integrity of the CORTEN steel test specimens coated with corrosion agent is quantified using acousto-ultrasonics. The time-frequency features of these acoustic waves are analysed in Mel scale. The acoustic waves propagated through the different configuration of the CORTEN steel specimens vary in their time frequency characteristics. These acoustic waves are classified using a Neural Network. An image classification based convolutional neural network (CNN) is built and trained for classifying the acoustic waves. 2. Materials and Methods 2.1. Materials Three different CORTEN steel specimens are tested in this study, which are named as Specimens A, B and C, respectively. The geometrical configurations of the CORTEN steel specimens are presented in Table 1. As presented in Table 1, all three specimens have same length and breadth. However, Specimen A is thin (thickness 0.8 mm), while Specimen B and C are relatively thick (thickness 2 mm). The surfaces of the specimens A and B are degreased and an oxide activator is coated on one side of both the specimens. The accelerated corrosion by the oxide activator is allowed to occur for 24 h. Post the stable oxide formation on the surface, the specimens are tested. Specimen C is tested under virgin conditions. The idea is to understand the attenuation of acoustic waves between Specimens A and B in terms of their geometrical differences and between specimens B and C in terms of the corrosion formation. Thereby, the attenuation of the acoustic waves due to geometrical differences and corrosion formation can be characterized separately.

Table 1. Geometrical Configurations of the Test Specimens

Thickness

Length

Breadth

Specimen Name

mm

mm 100 100 100

Mm 100 100 100

A B C

0.8

2 2

2.2. Acousto-Ultrasonic Test Setup For simulating the acoustic waves, a piezoelectric transducer (R30α Physical Acoustics) is spiked with a burst voltage of 28 V. This induces the stress waves, which reflect back from the surface. The first structural response is considered as the input reference. The acoustic waves are allowed to propagate for a distance of 30 mm and recorded using another R30α sensor. The signal from the transducer is amplified by 40 dB using 2 /4/6 AE preamplifer and recorded at a sampling rate of 1 MHz. Figure 1 shows the configuration of the test setup. 2.3. Convolutional Neural Network An image classification based Convolutional Neural Network (CNN) is built for classifying the acoustic waves generated during the test. It consists of 3 convolutional layers, each of which is followed by a pooling layer and ReLu activation function. Maximum pooling layer follows the convolutional layer 1 and average pooling layer follows the convolutional layer 2 and 3. It consists of a total of 9 hidden layers. For training the CNN, the stochastic gradient descent algorithm is used (Meng et. al., 2017). The CNN is trained with a piecewise learning, a learning rate of 0.001 and learn drop factor of 0.1. The architecture of the CNN is presented in Figure 2.