PSI - Issue 39

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

620

13

Finally, a KNN classifier tested the best indications. Zhao et al. (2020) presented a defect detection approach based on multi-input convolutional neural network (MI CNN) to address the difficulty in identifying shaft misalignment and crack in a rotor system. The MI-CNN was able to extract hidden information from raw vibration signals, with a recognition rate of 99.42 percent for health, shaft misalignment, crack, and misalignment-crack coupling of the rotor system, whereas the KNN, SVM, and RF struggled to distinguish different types of fault signals accurately. Söffker et al. (2015) compared a model-based and a signal based approach for crack detection and prediction in rotating machinery. SVM and wavelets as recent machine learning algorithms were used to introduce a fresh signal-based methodology. Modern machine learning techniques were shown to be more resistant to disturbances and sounds, but model-based strategies were found to be more flexible to system load changes and better able to interact with system physics and modelling parameters. Choudhury et al. (2019) compared DT regressor and RF regressor in order to anticipate crack parameters like position and depth before beam failure. The study used both a theoretical and numerical approach. The study included experimental analyses between 0°C and 50°C (the ambiance temperature). The results showed a 4-percentage-point disparity. The environment, noise, and so on all contributed to this error. The expected parameters might be achieved by establishing the vibration-fracture connection. The link produced beam natural frequencies, which were employed in regression techniques. The RF regressor predicted parameters more accurately than the DT regressor. Choudhury et al. (2019) considered two approaches considered: Least Absolute Shrinkage and Selection Operator (LASSO) for feature selection and regularisation and Ridge for collinearity issues. Both statistical models did well in detecting crack depth and location. The first three natural frequencies were independent on a cantilever beam structure, however the fracture location and depth were dependent. The derived regression model’s Bias Variance Trade-off was studied. (Bias is overfitting; variance is generalisation.) In terms of accuracy, ridge analysis outperformed LASSO. The theoretical results were compared to the experimental data, yielding a 4% inaccuracy. Liu and Meng (2005) examined the performance of neural network and SVM based on natural frequencies in crack identification. A supported beam with various types of cracks was employed in the experiment. Support vector regression theory and back-propagation neural networks were discussed. The possibility of crack identification using these methods was investigated by finding and sizing cracks in supported beams for which a few natural frequencies were available. It was discovered that crack position and depth might be approximated with a slight size error. The results showed that the SVM was an excellent and powerful fracture detection method. 4.8. Notable Mentions Behera et al. (2019) used a hybrid methodology to overcome the limitations of single AI systems in the detection, localisation, and severity of cracks. The experiment was conducted in the presence and absence of fractures on a fix hinge aluminium beam of the required dimensions. The hybrid intelligent model's outputs were compared to numerical and experimental data for fracture locations and depths. The article highlighted work that had been done to validate the correctness of hybrid controllers in a fix-hinge beam, which was unusual in the literature at the time. When the output findings for fracture location and crack depth were compared to experimental data, the deviation was found to be within acceptable bounds. Mohammad et al. (2014) used a multilayer perceptron (MLP) based system to detect rotor shaft defects. The Peak Position Component (PPC) was able to determine the aspects of each fault in the case of the specific scenarios. An ANN was fed Power Spectral Density (PSD) to detect cracks based on changes in the spectral content of the system’s vibration. The Peak Position Component Method (PPCM) reduced data transfer by statistically characterizing peak position in the PSD. PSD peaks were employed as a small fraction of the overall frequency range. The ANN used in the PPCM results was supervised feed-forward using Levenberg-Marquardt back propagation. Three networks were used, each with its own inputs. All networks with varying input sizes performed well. The results showed that the suggested approach may be applied successfully and reliably in real time. Outa et al. (2020) introduced a new technical tool, artificial immune systems (AISs). Computer tools for structural health monitoring include biological immune systems and AIS SHM. The experiment was separated into two phases: parametric characterization and standard characterization, which involved building a standard signal database and inserting two masses for frequency acquisition. This work developed an algorithm based on artificial AISs to recognize and classify signals as normal or faulty. The negative selection algorithm classified failures based on likelihood of failure and defect severity. The proposed SHM was found to be effective and dependable in prognosis and failure