NEW EXCITATION (MULTIPLE WIDTH PULSE EXCITATION (MWPE)) METHOD FOR SHM SYSTEMS—PART 2: CLASSIFICATION OF TIME- FREQUENCY DOMAIN CHARACTERISTICS WITH 2DSSD AND CNN

Surface response to excitation (SuRE) and electromechanical impedance methods quantify the difference between the reference and any given spectrums by calculating the sum of the squares of differences (SSD). In part one of this study, twodimensional SSD (2D-SSD) was proposed to quantify the difference of timefrequency plots when the part was excited with the Multiple Width Pulse Excitation (MWPE) signal. In this study, neural networks and deep learning were used for the classification of structural health monitoring (SHM) signals. Since manual encoding of the 2D spectrograms is very complicated to prepare them for classification by using neural networks, deep learning has been used. In this study, the performance of deep learning was evaluated for the classification of sensory data. A cross-shaped part made of PLA was manufactured additively and the center of the part was excited with MWPE and the surface waves were monitored at the end of each extension. Tests were repeated without and with a compressive force at each extension. The recorded time-domain sensory data was converted to spectrogram images using Short-Time Fourier Transform (STFT). The spectrograms were classified with the Convolutional Neural Network (CNN) after proper training. The results showed that the hidden geometry of each extension had a distinctive effect on the characteristics of the monitored signals. CNN could classify the infill type, skin thickness, and loading conditions with better than 92 % accuracy when the responses of the 20 pulses in the MWPE signal were considered.