PSI - Issue 39

Nabam Teyi et al. / Procedia Structural Integrity 39 (2022) 608–623 Author name / StructuralIntegrity Procedia 00 (2019) 000–000

619

12

Fig. 12. A random forest representation.

Fig. 13. A support vector machine representation. Thada et al. (2021) performed ML-based frequency modelling for structural crack detection, compared the performance of ML models and their ensembles in predicting the first six natural frequencies of a fixed terminated cracked beam. Based on the comparison of ML algorithms, they concluded that Neural Networks, Gradient Boosted DTs, and Extreme Gradient Boosted (XGB) Decision Trees had the best performance across three separate assessment metrics. Ensemble learning beat single estimator solutions. Hein and Jaanuska (2019) used Haar wavelet discrete transform, ANNs and RFs to anticipate the position and severity of a crack in an Euler–Bernoulli cantilever subjected to transverse free vibration. The results showed that the ensemble of 50 ANN trained using Bayesian regularization and Levenberg–Marquardt methods outperformed RF by a little margin. It was also shown that crack depth was more difficult to estimate precisely than crack location. Rodrigues et al. (2020) processed the spectrum image of vibration orbits obtained during rotating machine run-up to extract fault features such as imbalance, misalignment, shaft crack, rotor-stator rub, and hydrodynamic instability. The classifiers were trained on simulated data and then evaluated on both. The results showed that despite the high processing cost, the CNN technique performed the best for fault classification using spectral pictures. The Principal Component Analysis approach paired with KNN classifier had the lowest processing costs. Sánchez et al. (2020) employed condition-based monitoring to address the issue of early crack detection in railway axles. They did so by assessing several condition indicators of vibration signals in the time and frequency domains using two distinct methodologies. They first analysed solely the vibration signals obtained by accelerometers arranged longitudinally, and then utilised a data fusion technique to evaluate six accelerometers by merging the indicative conditions according to sensor location. In each case, 54 condition indicators were created for each vibration signal, with the best features selected using the Random Forest Mean Decrease Accuracy technique.