PSI - Issue 37

6

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

312

Fig. 6. Configuration of Acousto-Ultrasonic Test Setup

3.3. Identifying Corrosion Behaviour of CORTEN Steel using CNN

A brief summary on how the designed CNN can be used for identifying the corrosion behaviour in CNN is presented in this section. It is identified that the acoustic signal attenuation due to the stable corrosion formation is characterized by the loss in its power spectral density. This attenuation can be identified automatically using the trained CNN model. By simulating the stress waves or acoustic waves periodically through the CORTEN steel specimens, and recording and classifying them using CNN, the corrosion behaviour can be quantified. For example, in the present case, when the corrosion starts to occur, due to the loss in the power spectral density, the class of the propagating acoustic signal changes from Class B to Class C. This can be identified by the CNN, thereby identifying the corrosion progression. Nonetheless, a dedicated analysis on the different classes of acoustic signal propagating during different stages of corrosion progression must be made. In that way, the different stages of corrosion can be quantified using the CNN. Regardless, the proposed methodology in this research work is able to identify the attenuation of the acoustic signal due to the corrosion formation. 4. Conclusion The corrosion behaviour of the CORTEN steel specimens were studied using Acousto-Ultrasonics with the aid of a deep learning Convolutional Neural Network. First, it was identified that for an input signal of single frequency at 350 kHz, the propagating signals attenuate due to the geometrical variations and the corrosion formation. The attenuation due to the geometrical variations is in both the frequency and power spectral density characteristics. However, the attenuation due to the corrosion formation is strictly bound to the power spectral density characteristics. The power spectral density of the signal propagated in the specimen without corrosion is about 5 times larger than the power spectral density of the signal propagated in the specimen with corrosion. This attenuation is taken as the primary criteria for classifying the signals into three different classes: signals attenuated due to corrosion, geometrical variations and signals propagated when there is no corrosion occurrence. A Convolutional Neural Network (CNN) is designed and trained for classifying these signals. The efficiency of the trained CNN is more than 99%, proving its efficacy in classifying these signals, thereby identifying the class of signals propagated in the presence of corrosion.