PSI - Issue 39

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

613

6

act as if they were interconnected brain cells, results in ANNs. As in the biological brain, each link can communicate with other neurons. Signals from artificial neurons are processed and sent to other neurons. The output of each neuron is a non-linear function of its inputs. The edges join. Learning alters the weight of neurons and edges. A connection’s weight affects its signal strength. A neuron may only emit a signal if the aggregate signal passes a threshold. Bias is a constant that helps the model fit the data. Layers of neurons are common. Layers can alter their inputs in many ways. Signals cross numerous layers from input to output. With their training values, ANNs can model nonlinear situations and predict the output values of input parameters.

Fig. 6. An ANN architecture. A dynamic analysis of a shaft with two transverse cracks was performed by Saridakis et al. (2007), who used three parameters: position, depth, and relative angle. Each of the cracks was thought to be at a random angle to the longitudinal axis of the shaft and at a distance from the clamped end. Using fuzzy logic to determine crack features, two efficient goal functions were proposed and tested. An ANN approach to the analytical model was used with a genetic method to discover the fracture properties. To cut down on calculation time while still maintaining accuracy, researchers used evolutionary algorithms as well as neural network architectures. Cracks could be detected quickly, according to the results of the study. Techniques for detecting and anticipating multiple transverse cracks on a stepped rotor shaft were described by Baviskar & Tungikar (2013) and Baviskar & Tungikar (2014). Both cracks travelled parallel to the axis at first. Both cracks eventually grew vertical and slanted. The natural frequency was determined via modal analysis. ANN predicted fracture features in the forward technique. Using the natural frequency database, an ANN network was trained to predict crack properties. The applicability of the strategy was demonstrated by comparing ANN predictions to FEM and experiments. An FFT analyzer was used to test a stepped rotor shaft and a cantilever circular beam with two cracks. Kankar et al. (2012) used WEKA software to train ANN/SVM for classification of faults. Their methodology included a filtering algorithm that used a density-based clusterer to generate cluster membership values, as well as a filtering algorithm that selected the most appropriate characteristics. The effects of various vibration signals obtained with healthy and damaged rotors at varying speeds were studied. ANN had a classification accuracy of 95.54 %, which was somewhat higher than SVM at 94.43%. Sahin and Shenoi (2003) described an experimentally validated damage detection system in a beam like structure that used vibration-based analytical data as input for ANNs . During training, ANNs were subjected to simulated vibration responses with and without artificial noise. Before introducing extracted characteristics to ANNs, they were subjected to sensitivity analysis using various vibration modes, taking into account damage location and severity. Given the ANNs’ tolerance for increased noise, localization was more accurate than quantification. The level of noise on the features, the experimental procedure, the measuring equipment, and their usefulness in predicting the severity and location of damage in beamlike structures all played a role. Sahin and Shenoi (2003) again studied vibration based properties of a beam-like composite laminate to forecast damage severity and location using feedforward back propagation neural networks. Combining three criteria, namely reduction in natural frequencies, maximum absolute changes in curvature mode forms and their placements, resulted in less promising results for damage severity and location. In the absence of noise, two untrained ANNs outperformed one trained ANN. However, when given more noise-polluted data, ANNs provided more accurate and robust damage localization predictions than damage quantification. A study of multiple ANNs trained on vibration-based analytical data revealed that the features retrieved and used as inputs, as well as the noise on these features, influence the accuracy of ANNs’ structural damage