PSI - Issue 64

Ricardo Perera et al. / Procedia Structural Integrity 64 (2024) 1369–1375 Author name / Structural Integrity Procedia 00 (2019) 000–000

1372

4

Fig. 3. Experimental set-up.

Fig. 4. Photo of the experimental set-up and detail of the sensors

EMI signals were collected from the eight sensors by an impedance analyzer with a sampling frequency of 12.5 Hz within a range between 10 kHz and 100 kHz, under different damage conditions of the beam. To avoid uncertainties in the experimental results, each experiment was repeated five times to obtain the average value. The data collected under healthy conditions (baseline dataset) are regarded as normal samples, while the data collected under five damaged conditions are regarded as unhealthy samples. Baseline dataset was used to train the convolutional autoencoder. 4. Results and discussion This section analyzes the performance of the proposed deep autoencoder model based on the measured (original) and reconstructed EMI spectrums. This section also describes the damage-sensitive feature of the proposed deep autoencoder model, corresponding to different beam conditions (undamaged and damaged). Firstly, Fig. 5 shows the impedance vs frequency (Hz) and compares the measured (original) and reconstructed EMI spectrums from the undamaged baseline dataset, used to train the network, for sensors 3, 4, 7 and 8 in Fig. 3. It shows that the proposed deep autoencoder can accurately reconstruct the EMI spectra since the difference between both original and reconstructed spectrums is minimum.