Analysis of noise effects with Deep Learning and Structural Health Monitoring applications

Structural Health Monitoring (SHM) has always been a hot topic in the field of Mechanical and civil engineering, and its first-level task is damage detection. Traditional Stochastic Subspace Identification (SSI) performs damage detection through operational modal analysis (OMA), which is a data-driven method and has high identification accuracy, but huge amount of data. Advances in artificial intelligence tools represent the frontier of SHM, enabling non-destructive assessments directly based on the output-only vibration signals. In this thesis, three methods are proposed to deal with multi-class damage classification tasks, the first method relies only on statistical features, in contrast, the second method considers the most informative subspace-based damage indicator. Finally, the third method considers the entire set of damage indicators obtained through a reasonably variable range of input parameters as features. Two Deep Learning tools – Multi-layer Perceptron (MLP) and 1D-convolutional network (CNN) are used to compare and verify the results. To verify the noise effect on the SSI method and Deep Learning model, three groups of different Signal-to-Noise Ratio (SNR) levels were set up to simulate sensors in real working condition. The proposed experiments are tested on numerical benchmark problems. The results show that, even in the data collected by real sensors, Deep Learning models and SSI methods have good noise immunity. Subspace-based damage-sensitive features is still the most effective indicator in the field of Structure health monitoring