Experimental validation of model-free damage detection approaches based on using modal features

Abstract

Abstract

Structural Health Monitoring involves identifying and assessing structural changes or damage with or without relying on predefined numerical models. Model-free methods exploit only damage-sensitive dynamic characteristics obtained directly from measured responses. Some of these are based on hypotheses regarding expected damage, which turn into recognizable footprints. Alternatively, model-free methods require the availability of a baseline that refers to an intact condition. The proposed damage identification methods fall into the latter category because they enable efficient and reliable monitoring of structural integrity under operational conditions using an array of accelerometers. The scope of this study is to validate both methods experimentally by applying them to scale-model tests simulating the onset of damage on a real ship. The first method is based on a macro index, mapping the probability of damage over the investigated structure. It incorporates several damage indices that process the sampled modal curvatures. These modal curvatures, linearly averaged or squared over the sensor mesh (geometric strain energy), are computed in both the reference-intact and current-damage conditions and fed into the different functions characterizing the various indices. If damage is effectively present in one element of the sensor mesh, damage severity can be inferred from the index value. Nonetheless, damage localization can be improved if the indices are Z-score normalized and combined into the macro-index using ensembling strategies. Thus, thresholds for damage existence can be properly set low on the index average to have sufficient sensitivity, while agreement conditions among indices can be exploited to reduce false warnings. The second method is a novelty detection approach based on a histogram score. It shares with the previous method also the use of one of the indices considered above (the Cornwell’s formulation of the Modal Strain Energy Index) though in a different way. The operational modal analysis provides the vibration modes from the tests, which are characterized by noise from different sources. To train the method with a sufficiently large ‘intact’ population, the mode shapes are first averaged and then contaminated with Gausisan noise, experimentally modelled. The statistical distribution of the Cornwell’s damage index provides the baseline to evaluate whether the same quantitity, computed directly from the experimental data, can be attributed to an underlying structural modification. The threshold to separate the intact and damaged classes plays again a crucial role. The damage identification techniques are validated using data collected in scale-model tests of a navy vessel within the “Digital Ship Structural Health Monitoring project” (dTHOR), granted by the European Defence Fund, aimed at developing a system based on innovative utilization of extensive on-board measurements, a comprehensive digital framework, and hybrid analysis and modelling. The ship longitudinal bending stiffness is reproduced by an elastic backbone connecting the hull portions. The damage is artificially generated by removing the plate elements on top and side faces of the aluminium backbone and affects the operational modes identified while the physical model is towed at the CNR-INM wave basin. Results, though quite promising, are critically reviewed in the perspective of further increasing their accuracy.