The basic problem of damage detection is to deduce the existence of a defect in a structure from measurements taken at sensors distributed on the structure. Especially in aeronautical structures, cracks, delamination and debonding are typical types of damages often encountered. The problem is essentially one of pattern recognition. Artificial neural networks show considerable promise for damage diagnosis.
In the most basic, supervised learning, approach to deriving a neural network, the network is presented with pairs of data vectors, the input being the vector of measurements from the system and the output being the desired damage classification. At each presentation of the data, the internal structure of the network is modified, in order to bring the actual network outputs into correspondence with the desired outputs. This iterative procedure is terminated when the network outputs have the required properties over the whole training set. In a structural application, the training data may be provided by finite element (FE) analysis. This has the advantage of allowing a large range of boundary conditions and static/dynamic load cases to be analysed. FE analysis may be a little unrealistic as there is no limit on the spatial resolution of the data which is obtained, e.g. strains. In reality, the number of sensors available will be limited and this will, of course, place restrictions on the resolution of data. As a result, it is necessary in practice to optimise the number and location of sensors for a given problem.
The main objective of the current proposal is to develop a mathematical algorithm for optimal strain sensor (strain gauges, fibre Bragg grating or other) placement in aeronautical composite structures for maximum damage detectability. The mathematical method to be used will be a genetic algorithm based on neural networks. The genetic algorithm will be trained from finite element analyses simulating impact scenarios (damage initiation) and operational