PSI - Issue 77

Claudia Barile et al. / Procedia Structural Integrity 77 (2026) 3–10 Barile and Kannan/ Structural Integrity Procedia 00 (2026) 000 – 000

5 3

Fig. 1 - Experimental Flexural Test Setup.

The AE signals are acquired using a pair of piezoelectric sensors. The piezoelectric sensors, PICO, are placed approximately 50 mm from the midspan of the specimen and held firmly using a mechanical clip. PICO sensor acquires AE signals with high accuracy within 200 to 750 kHz frequency bands. Therefore, the signals are acquired at a sample rate of 2 MHz as per Nyquist sampling theorem. AE signals crossing the detection threshold of 35 dB are acquired, amplified by 40 dB using a preamplifier and filtered through a 100 kHz/600 kHz bandpass filter. The filtered signals are recorded in burst-mode at 2 MHz sample rate and 1024 sample length. 2.3. Microscopic Analysis Fractured surfaces of the tested specimens are analysed under optical microscope NIKON SMZ800. It is a white light microscope with a maximum magnification of 6.3x. The surfaces are cleaned with isopropanol and are analysed. 2.4. Machine Learning-based Acoustic Emission Data Analysis The AE signals from the different specimens during the flexural tests are acquired using the piezoelectric transducers. The significant signal features suitable for identifying the damage progression are extracted from the AE signal data. In literature, the Weighted Peak Frequency (WPF) and Peak Amplitude (PA) are used to associate the AE signals with the damage sources (Barile et al., 2020; De Groot et al., 1995; Muir et al., 2021). The WPF and PA of the signals are used to label and classify the AE data in an unsupervised manner using k means++ data clustering algorithm. Before classifying the data, Davies-Bouldin Index (DBI) is used to evaluate the optimal number of clusters present in the AE data distribution. Following this, the AE signals are classified into the predefined number of clusters. 2.5. Signal Analysis using Continuous Wavelet Transform The AE signals from different clusters are analysed in the Time-Frequency (TF) domain using CWT. The AE signal waveforms of length 1024 samples are decomposed into TF coefficients using ‘analytical Morlet’ wavelet function in the CWT. CWT results provide the time-frequency distribution of the signal content as spectrograms (Krishna et al., 2025). The spectrograms of AE signals from different clusters are further analysed and compared with each other.