On the complexity of artificial neural networks for smart structures monitoring

A Bayesian probabilistic approach is presented for smart structures monitoring (damage detection) based on the pattern matching approach utilizing dynamic data. Artificial neural networks (ANNs) are employed as tools for matching the “damage patterns” for the purpose of detecting damage locations and estimating their severity. It is obvious that the selection of the class of feed-forward ANN models, i.e., the decision of the number of hidden layers and the number of hidden neurons in each hidden layer, has crucial effects on the training of ANNs as well as the performance of the trained ANNs. This paper presents a Bayesian probabilistic method to select the ANN model class with suitable complexity, which is usually overlooked in the literature. An example using a five-story building is used to demonstrate the proposed methodology, which consists of a two-phase damage detection method and a Bayesian ANN design method.