Estimating Confidence in Damage Position Predictions Made Involving ANN

The evaluation of structural damage using global methods has become a common engineering practice. The damage location and extent are related to the change in modal parameters; if the modal parameter changes are known, damage assessment becomes an inverse problem. In this study, we use natural frequency changes (RFS) as damage indicators and an artificial neural network (ANN) to make the association with the damage location and extent. Previously, we derived a mathematical relation to express the RFSs for bending vibration modes with respect to the crack parameters. After a second normalization, we obtain damage location coefficients (DLC), that are related just to the crack location. We use this relationship and create a database to train the ANN. The RFSs and separately the DLCs constitute input data, and the crack location is the output data. Several networks with different hyperparameters are trained and tested against RFSs obtained from numerical simulations. The estimates are fairly accurate, but each network estimated the crack to be in a different location. To make a correct decision, the decision-maker must be aware of the confidence he can have in the predicted results. Therefore, we introduced explainability into the ANN responses, by inserting an additional result, namely the recalculated DLCs for the estimated crack position. By comparing this set of DLCs with the one resulting from the measurements, we can deduce if the area identified as containing the crack is the correct one or if another ANN model should be used.