PSI - Issue 37

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

311

5

frequency. However, how this may quantify the corrosion characteristics of these Specimens? The signals propagated in Specimens B and C, although having same frequency characteristics, have large variations in terms of power spectral density. The spectral density of the signal recorded in Specimen B is about 5 times less than the signal recorded in Specimen C. It must be noted that the corrosion is formed on the surface of Specimen B. Typically, most of the energies of the acoustic wave propagated over the surface of the specimens. Since the corrosion in Specimen B induces surface roughness, there is a huge energy loss during the acoustic wave propagation. This leads to the loss in power spectral density. Therefore, by quantifying the power spectral density between the non-corroded and corroded specimens, a quantitative extent of surface corrosion can be identified. Although this requires further deep investigation, the acousto-ultrasonic results are promising. 3.2. Convolutional Neural Network Results Since the frequency characteristics of the signal propagated through the different specimens are quite different, the occurrence of corrosion can be identified. By continuously monitoring the propagating acoustic signals, by comparing the changes in the power spectral density, the initiation of corrosion and its stabilization can be achieved. To automate this process, in this research work, a deep CNN is used. 10000 waveforms propagated through each of the specimen are taken as the training data (a total of 30000 training data) and 5000 waveforms are taken as the test data. The image based classification extracts the features from the training data and trains the CNN. Typically, time series representations of the signal or their time-frequency characteristic representations such as Short-Time Fourier Transform (STFT), Continuous Wavelet Transform (CWT) or other spectrograms are used. Since the signals recorded in this study is very transient and localized to a very narrow frequency band, Mel spectrograms are used. Mel spectrograms rescales the signals to their Mel frequency equivalents and provides the results in the time-frequency domain. As an example, Mel spectrograms of a signal recorded in Specimens A, B and C each are presented in Figure 5. Besides, Mel spectrograms are most commonly used as the input image form in training neural networks. The signals recorded from Specimens A, B and C, respectively are classified as Classes A, B and C. Class A represents the acoustic signals attenuated due to geometric variations, Class B the acoustic signals attenuated due to the corrosion formation and Class C the signals propagated when there is no corrosion.

Fig. 5. Mel Spectrograms of the signals propagated in (a) Specimen A (b) Specimen B and (c) Specimen C

The confusion matrix showing the results of the test dataset is presented in Figure 6. The results show that the efficiency of the CNN is more than 99%. In fact, only one signal in Class B is misidentified as Class C, while the rest of the signals are classified efficiently. This shows that the designed CNN is more than efficient in classifying the different classes of acoustic signals.