|
Health assessment of structures is an essential and innovative method to evaluate the behavior of the structure and to estimate the performance during the service period. Health assessment is important not only for structural safety and integrity but also for human safety requirements. A discontinuity in the form of damage changes the global structural stiffness, which in turn modifies the modal response of the structure. This assimilation can be achieved from changes in the structure's natural frequencies, one of the modal parameters that can change globally. In this present investigation, modeling of different cantilever beams having different levels of damage severities is performed. Finite Element (FE) simulations is performed using 3D modal analysis methodology for different material properties. For this, eight noded brick elements is used with fine mesh size. A dataset is created in the form of an array with natural frequencies for different cases. In recent years, neural networks have shown promising results in applications like image and text classification, machine translation, anomaly detection, and many more. In this work, modal analysis is used along with Artificial Neural Networks (ANN) for damage severity assessment. We have utilized ANN for classifying the modal signatures into four different damage severity classes. The dataset is used to train a classification-based neural network, categorizing natural frequencies (inputs) into damage severity classes (outputs). We have used a categorical cross-entropy loss function with the Adam optimization scheme. We have seen that the networks are able to train with low loss and higher accuracy. The performance of the network is measured using metrics like classification accuracy, F1-score, confusion matrix and Receiver Operating Characteristic (ROC) curves, training, and prediction time. New signatures are fed into the trained networks to identify the damage levels autonomously without waiting for a time-taking expert's analysis procedure. It is seen that the network can generalize well on unseen examples and can predict the health of the structure.
|
