Genetic optimisation of a neural damage locator

A critical problem in structural health monitoring (SHM) based on pattern recognition methods is the correct selection of features, i.e. measured and processed data for the diagnosis. Various selection strategies have been applied in the past and one approach that has proved effective is the use of combinatorial optimisation methods. This paper presents a case study based on a scheme for damage location in an aircraft wing. The feature selection algorithm is a Genetic Algorithm and the locator (classifier) is an artificial neural network. A comparison is made with the results obtained when the features are selected on the basis of engineering judgement. The study is seen to raise some issues relating to model complexity and generalisation and these matters are discussed in some detail.