Subspace features and statistical indicators for neural network-based damage detection

The late opportunities prompted by artificial intelligence have motivated the current research about structural damage detection strategies based on damage-sensitive subspace-based indicators (DI). Precisely, three different methodologies (A), (B), and (C) are discussed for multiclass damage classification with a multi-layer perceptron (MLP) network. Specifically, the network's inputs combine vibration response statistics with subspace-based features. Method (A) relies on statistical features only, whereas method (B) also considers the most informative subspace-based DI, retrieved from an empirical sensitivity analysis. Finally, method (C) provides a new perspective by overcoming the arbitrary choice of parameters affecting the subspace-based DIs computation. These three methods are tested on a numerical benchmark problem, and the results emphasize the last approach as the most promising methodology. For the sake of further validation purposes, the three methods have been finally tested on an experimental steel I-beam setup, evidencing the effectiveness of informative subspace-based DIs.