Problem specific neural networks (NN) are able to locate and quantify damages in plane truss structures due to their ability to learn unknown functional relationships and adjust to complex behavior. First, the behavior of the undamaged structure as well as the behavior of the structure including various states of damage are calculated or obtained from experimental results. The training patterns describe the characteristic behaviour of the structure. It is shown that a trained neural network is able to detect and quantify existing damages in the structural system, based on the range of the presented training data. This holds under the assumption that the training patterns include a sufficient number of characteristics that distinguish between distinct states of the structural behavior. The neural network therefore approximates an inverse functional relation, which is virtually unsolvable using conventional analytical methods. This paper evaluates an example in the field of a two dimensional statically indeterminate truss structure, reduction (pruning) of a fully connected initial neural network topology and a particular problem-specific data preprocessing (dimensional analysis). A statically indeterminate system is investigated with a main focus on sufficient data diversification and adequate data preprocessing. Using sensitive pruning algorithms (Optimal Brain Surgeon), a general initial network topology is minimized to a state which is necessary to comprehensively describe the given problem and thereby assess the impairment of a damaged structural system with a minimized neural network topology.
展开▼
