MS17.2 - Structural Health Monitoring

An autonomous Structural Health Monitoring (SHM) strategy is presented for the online identification and classification of different types of Acoustic Emission (AE) sources in composite structures. Towards this, a Convolutional Neural Network based deep learning network is prepared for the automatic detection and characterization of pencil-lead break damage (PLBD)- resembling artificial debonding generation, low-speed tool-drop (TD) and steel ball-drop (BD) impacts on a laboratory-based carbon-fibre reinforced composite structure. The proposed deep network uses the AE signals corresponding to a series of PLBD, TD and BD events on different predefined zones of the targeted composite structure. In the process, the registered AE signals time domain-domain are converted to time-frequency scalogram by performing the continuous wavelet transform. These scalogram images are then resized to reduce the computational effort. The classification results show high accuracy in the training, validation, and testing of the network. The study was further extended for the AE signals under environmental impacts. The results show good accuracy under variable environmental conditions. It is envisaged that this deep learning based autonomous SHM strategy can be translated for the real-time monitoring of several other structural components under variable operating conditions and applications.

Population-based Structural Health Monitoring (PBSHM) aims to gain additional insights on the health of a structure when using data available across a population of similar structures, as compared to the insight available when using only data from a single structure. Before knowledge can be transferred across structures, the similarity between structures (or substructures), within the population must be established. The first paper in the series explored the use of Graph Neural Networks (GNNs) to compute similarity measures via an Irreducible Element (IE) model representation of structures stored within a PBSHM database. The second paper in the series explored the use of a comparison layer, using the Canonical Form representations of the associated structures to compare against. This paper builds upon the aforementioned research and explores the viability of using clustering algorithms to determine which population a structure belongs to within the network of structures from the CF similarity score.

Damage detection is the monitoring of the onset of damage in the structural system response, deviating from a reference undamaged state. The application of data-based Structural Health Monitoring (SHM) for damage detection is characterized by three fundamental aspects: the features extracted and selected as representative of damage in the structure, the metrics used as novelty or damage index and the model built to highlight underlying patterns indicative of the presence of damage. Focusing on the first step towards the optimal application of the data-based SHM approach, the extracted parameters should be truly sensitive to the presence of damage, robust to noise and distinguishable from the environmental and operational variability. Great research effort has been spent into the study of the sensitivity to damage of modal parameters. Numerous publications are available reporting requirements, potential and limits of parameters such as resonance frequencies, mode shapes, mode shapes curvature, modal flexibility, modal strain energy, etc. On the other hand, the sensitivity to damage of statistical parameters have also been investigated with the use of autoregressive models, performing damage detection directly on vibration data and avoiding the modal identification of the structure. This paper addresses a comparison between the use of modal and statistical parameters as damage sensitive features in a damage detection problem. This problem is approached as an outlier detection problem and the features used are evaluated based on the detection performance, the extraction/selection requirements, and the computational cost. In order to control all parameters that could affect the structural response, an artificial dataset is generated from the numerical model of a three-storey steel frame structure in different damage scenarios.