Potential and limitations of NARX for defect detection in guided wave signals

Previously, a nonlinear autoregressive network with exogenous input (NARX) demonstrated an excellent performance, far outperforming established method in optimal baseline subtraction, for defect detection guided wave signals. The principle is to train NARX on defect-free signals obtain filter that predicts the next point from previous points signal. trained then applied new measurement and output subtracted reveal presence of responses. However, as shown this paper, performance implementation lacks robustness; it highly dependent initialisation sometimes improves worsens over course training. It due only making predictions one ahead. Subsequently, multi-step prediction using newly proposed structure creates more robust training procedure, by enhancing correlation between loss metric performance. physical significance explored, allowing simple hyperparameter tuning strategy be used determining structure. overall also improved prediction, responses at different times well data sensor pairs, revealing generalisability method.