Damage detection on crates of beverages by artificial neural networks trained with finite-element data Original Research Article

Recognition of representative damages on returnable crates of beverages is carried out by an artificial neural network (ANN) trained exclusively with frequency response spectra from finite-element simulations. Finite-element mode shape analysis implies that two sensors are sufficient for successful damage recognition. Amongst three topologies, a loose coupling of two ANN yields best recognition results, each processing data from one sensor. Out of 91 experimental recordings 65 of 66 data sets representing twenty damage types are recognised. The classification fails for some data sets of intact crates, due to experimental conditions not accounted for in the finite-element simulation.