A novel information-gap technique to assess reliability of neural network-based damage detection

The application of neural network classifiers to a damage detection problem is discussed within a framework of an interval arithmetic-based information-gap technique. Using this approach the robustness of trained classifiers to uncertainty in their input data was assessed. Conventional network training using a regularised Maximum Likelihood approach is discussed and compared with interval propagation applied as a tool to evaluate the robustness of a particular network. Concepts of interval-based worst-case error and opportunity are introduced to facilitate the analysis. The interval-based approach is further developed into a network selection procedure capable of significant improvements (up to 22%) in the worst-case error performance over a conventional network trained on crisp (single-valued) data.