A Self-adaptive Hybrid Model/data-Driven Approach to SHM Based on Model Order Reduction and Deep Learning

Aging of structures and infrastructures urges new approaches to ensure higher safety levels without service interruptions. Structural health monitoring (SHM) aims to cope with this need by processing the data continuously acquired by pervasive sensor networks, handled as vibration recordings. Damage diagnosis of a structure consists of detecting, localizing, and quantifying any relevant state of damage. Deep learning (DL) can provide an effective framework for data processing, regression, and classification tasks used for the aforementioned damage diagnosis purposes. Within this framework, we propose an approach that exploits a deep convolutional neural network (NN) architecture. The training of the NN is carried out by exploiting a dataset, numerically built through a physics-based model of the structure to be monitored. Parametric model order reduction (MOR) techniques are then exploited to reduce the computational burden related to the dataset construction. Within the proposed approach, whenever a damage state is detected, the physical model of the structure is adaptively updated, and the dataset is enriched to retrain the NN, allowing for the previously detected damage state as the new baseline.