Structural health monitoring for combined damage states

The current state of the art in automated structural health monitoring (SHM) is to compare the results of a set of sensors to a set of previously established values, picking the most similar set. To identify a damage state, it must be envisioned upon the creation of the SHM system and analysed to determine the output. Following catastrophic events, bridges may be damaged in several ways in several locations, causing problems for the automated system in identifying the exact repairs that are necessary or whether sufficient strength remains to sustain reduced loading. An alternative method to modelling a massive number of damage states and memorizing them all is presented based on subdivided attractor networks. These artificial neural networks are similar to the feedforward networks that have found much use in industry. Instead of acting as an approximator for some function, these networks work as content addressable memory. Partitioning these networks allows different portions of the network to represent different portions of the structure, and to mix and match states that may be associated with each other. In this way a SHM regimen is produced that needs far fewer memorized states to recall all of the possible damage states.